msc 2020 classi cation - people.math.ethz.ch
TRANSCRIPT
MODEL-FREE PORTFOLIO THEORY: A ROUGH PATH APPROACH
ANDREW L. ALLAN, CHRISTA CUCHIERO, CHONG LIU, AND DAVID J. PROMEL
Abstract. Based on a rough path foundation, we develop a model-free approach to sto-chastic portfolio theory (SPT). Our approach allows to handle significantly more generalportfolios compared to previous model-free approaches based on Follmer integration. With-out the assumption of any underlying probabilistic model, we prove a pathwise formula forthe relative wealth process which reduces in the special case of functionally generated portfo-lios to a pathwise version of the so-called master formula of classical SPT. We show that theappropriately scaled asymptotic growth rate of a far reaching generalization of Cover’s univer-sal portfolio based on controlled paths coincides with that of the best retrospectively chosenportfolio within this class. We provide several novel results concerning rough integration,and highlight the advantages of the rough path approach by considering (non-functionallygenerated) log-optimal portfolios in an ergodic Ito diffusion setting.
Key words: stochastic portfolio theory, Cover’s universal portfolio, log-optimal portfolio,model uncertainty, pathwise integration, rough path.MSC 2020 Classification: 91G10, 60L20.
1. Introduction
Classical approaches to portfolio theory, going back to the seminal work of H. Markowitz[Mar59] (see also the early work of B. de Finetti [dF40]), are essentially based on simplisticprobabilistic models for the asset returns or prices. As a first step classical portfolio selectionthus requires to build and statistically estimate a probabilistic model of the future assetreturns. The second step is usually to find an “optimal” portfolio with respect to the now fixedmodel. However, it is well known that the obtained optimal portfolios and their performanceare highly sensitive to model misspecifications and estimation errors; see e.g. [CZ93, DGU07].
In order to account for model misspecification and model risk, the concept of model am-biguity, also known as Knightian uncertainty, has gained increasing importance in portfoliotheory; see e.g. [PW07, GR13]. Here the rationale is to accomplish the portfolio selection withrespect to a pool of probabilistic models, rather than a specific one. This has been pushedfurther by adopting completely model-free (or pathwise) approaches, where the trajectoriesof the asset prices are assumed to be deterministic functions of time. That is, no statisticalproperties of the asset returns or prices are postulated; see e.g. [PW16, SSV18, CSW19].In portfolio theory there are two major approaches which provide such model-free ways ofdetermining “optimal” portfolios: universal and stochastic portfolio theory.
The objective of universal portfolio theory is to find general preference-free well performinginvestment strategies without referring to a probabilistic setting; see [LH14] for a survey. Thistheory was initiated by T. Cover [Cov91], who showed that a properly chosen “universal”portfolio has the same asymptotic growth rate as the best retrospectively chosen (constantlyrebalanced) portfolio in a discrete-time setting. Here, the word “universal” indicates themodel-free nature of the constructed portfolio.
Date: May 5, 2022.1
2 ALLAN, CUCHIERO, LIU, AND PROMEL
Stochastic portfolio theory (SPT), initiated by R. Fernholz [Fer99, Fer01], constitutes adescriptive theory aiming to construct and analyze portfolios using only properties of ob-servable market quantities; see [Fer02, KF09] for detailed introductions. While classical SPTstill relies on an underlying probabilistic model, its descriptive nature leads to essentiallymodel-free constructions of “optimal” portfolios.
A model-free treatment of universal and stochastic portfolio theory in continuous-time wasrecently introduced in [SSV18, CSW19], clarifying the model-free nature of these theories. Sofar this analysis has been limited to so-called (generalized) functionally generated portfolios,cf. [Fer99, Str14, SSV18]. These are investment strategies based on logarithmic gradients of so-called portfolio generating functions. This limitation is due to the fact that the correspondingportfolio wealth processes can be defined in a purely pathwise manner only for gradient-typestrategies, namely, via Follmer’s probability-free notion of Ito integration; see Follmer’s pio-neering work [Fol81] and its extensions [CF10, CP19]. Even though these limitations do notoccur in discrete time, optimal portfolio selection approaches based on functionally gener-ated portfolios have also gained attention in discrete time setups; see e.g. [CW21]. Anotherstrand of research is robust maximization of asymptotic growth within a pool of Markovianmodels as pursued in [KR12, KR21, IL22]. While these approaches clearly account for modeluncertainty, a probabilistic structure still enters via a Markovian volatility matrix and aninvariant measure for the market weights process. In a similar direction goes the constructionof optimal arbitrages under model uncertainty as pioneered in [FK11].
The main goal of the present article is to develop an entirely model-free portfolio theory incontinuous-time, in the spirit of stochastic and universal portfolio theory, which allows one towork with a significantly larger class of investment strategies and portfolios. For this purpose,we rely on the pathwise (rough) integration offered by rough path theory—as exhibited ine.g. [LQ02, LCL07, FV10, FH20]—and assume that the (deterministic) price trajectories onthe underlying financial market satisfy the so-called Property (RIE), as introduced in [PP16];see Section 2.2. While Property (RIE) does not require any probabilistic structure, it issatisfied, for instance, by the sample paths of semimartingale models fulfilling the conditionof “no unbounded profit with bounded risk” and, furthermore, it ensures that rough inte-grals are given as limits of suitable Riemann sums. This is essential in view of the financialinterpretation of the integral as the wealth process associated to a given portfolio.
In the spirit of stochastic portfolio theory, we are interested in the relative performanceof the wealth processes, where the word “relative” may be interpreted as “in compari-son with the market portfolio”. In other words, given d assets with associated price pro-cess S = (S1
t , . . . , Sdt )t∈[0,∞) satisfying Property (RIE), we choose the total market capi-
talization S1 + · · · + Sd as numeraire, so that the primary assets are the market weightsµ = (µ1
t , . . . , µdt )t∈[0,∞), given by
µit :=Sit
S1t + · · ·+ Sdt
, i = 1, . . . , d,
which take values in the open unit simplex ∆d+. The main contributions of the present work
may be summarized by the following.
• In Proposition 3.9 we establish a pathwise formula for the relative wealth processassociated to portfolios belonging to the space of controlled paths, as introducedin Definition 2.3 below. This includes functionally generated portfolios commonlyconsidered in SPT—as for instance in [Str14, SV16, KR17, RX19, KK20]—as well as
PORTFOLIO THEORY WITH ROUGH PATHS 3
the class which we refer to as functionally controlled portfolios, which are portfoliosof the form
(1.1) (πFt )i = µit
(F i(µt) + 1−
d∑j=1
µjtFj(µt)
),
for some F ∈ C2(∆d+;Rd). Here, (πF )i denotes the proportion of the current wealth in-
vested in asset i = 1, . . . , d. In the case of functionally generated portfolios, i.e. when Fis the logarithmic gradient of some real valued function, we also derive in Theorem 3.11a purely pathwise version of the classical master formula of SPT, cf. [Fer02, Str14].• We introduce Cover’s universal portfolio defined via a mixture portfolio based on
the notion of controlled paths, and show that its appropriately scaled logarithmicrelative wealth process converges in the long-run to that of the best retrospectivelychosen portfolio; see Theorems 4.9 and 4.12. This extends the results of [CSW19] toa considerably larger class of investment strategies.• We compare Cover’s universal portfolio with the log-optimal portfolio assuming an
ergodic Ito diffusion process for the market weights process. In this case the corre-sponding growth rates are asymptotically equivalent, as shown in Theorem 5.4.• We develop novel results in the theory of rough paths to allow for the pathwise treat-
ment of portfolio theory. In particular, these results include an extension of [PP16,Theorem 4.19], stating that the rough integral can be represented as a limit of left-point Riemann sums—see Theorem 2.12—and the associativity of rough integration,exhibited in Section A.2.
One important motivation for our work comes from classical considerations of the log-optimal portfolio in ergodic Ito diffusion models for the market weights process. Indeed, this isone prominent example of an “optimal” portfolio that does not belong, in general, to the classof (generalized) functionally generated portfolios, but is still a functionally controlled portfolioof the form (1.1); see Section 5.2. As illustrated numerically in Figure 1, the log-optimalportfolio (an example of a functionally controlled portfolio) might significantly outperforma corresponding “best” functionally generated portfolio. Indeed, the blue line illustrates theexpected utility of the log-optimal portfolio over time, whereas the orange line depicts thatof a certain best functionally generated portfolio. For the details of this example we refer toSection 5.3. This indicates that going beyond functionally generated portfolios can have asubstantial benefit. This holds true in particular for Cover’s universal portfolio when definedas a mixture of portfolios of the form (1.1), since in ergodic market models it asymptoticallyachieves the growth rate of the log-optimal portfolio (see Theorem 5.4). Note that, due to therough path approach, both the relative wealth processes obtained by investing according to thelog-optimal portfolio and according to the universal portfolio make sense for every individualprice trajectory. This also gives a theoretical justification for learning a (non-functionallygenerated) log-optimal portfolio from the observations of a single price path.
Outline: In Section 2 we provide an overview of the essential concepts of rough paths andrough integration relevant for our financial application. In Section 3 we introduce the pathwisedescription of the underlying financial market and study the growth of wealth processesrelative to that of the market portfolio, which leads us to a pathwise master formula analogousto that of classical SPT. Section 4 is dedicated to Cover’s universal portfolio and to provingthat its appropriately scaled asymptotic growth rate is equal to that of the best retrospectively
4 ALLAN, CUCHIERO, LIU, AND PROMEL
0 50 100 150 200 250
0
0.1
0.2
0.3
0.4
days
log-optimalalpha-optimal
Figure 1. Expected utility of the log-optimal vs. the alpha-optimal portfolioover time.
chosen portfolio. In Section 5 we compare Cover’s universal portfolio with the log-optimalone, assuming that the market weights process follows an ergodic Ito diffusion. In this setting,we also compare the wealth processes of functionally controlled portfolios and functionallygenerated ones, illustrating their performance by means of a concrete numerical example.Appendices A and B collect findings concerning rough path theory and rough integrationneeded to establish the aforementioned results.
Acknowledgment: A. L. Allan gratefully acknowledges financial support by the SwissNational Science Foundation via Project 200021 184647. C. Cuchiero gratefully acknowl-edges financial support from the Vienna Science and Technology Fund (WWTF) under grantMA16-021 and by the Austrian Science Fund (FWF) through grant Y 1235 of the START-program. C. Liu gratefully acknowledges support from the Early Postdoc. Mobility Fellowship(No. P2EZP2 188068) of the Swiss National Science Foundation, and from the G. H. HardyJunior Research Fellowship in Mathematics awarded by New College, Oxford.
2. Rough integration for financial applications
In this section we provide the essential concepts from rough path theory for our applicationsin model-free portfolio theory. Additional results regarding rough integration are developedin the appendices. For more detailed introductions to rough path theory we refer to the books[LQ02, LCL07, FV10, FH20]. Let us begin by introducing some basic notation commonlyused in the theory of rough paths.
2.1. Basic notation. Let (Rd, | · |) be standard Euclidean space and let A ⊗ B denote thetensor product of two vectors A,B ∈ Rd, i.e. the d× d-matrix with (i, j)-component given by[A ⊗ B]ij = AiBj for 1 ≤ i, j ≤ d. The space of continuous paths S: [0, T ] → Rd is given byC([0, T ];Rd), and ‖S‖∞,[0,T ] denotes the supremum norm of S over the interval [0, T ]. For
the increment of a path S: [0, T ]→ Rd, we use the standard shorthand notation
Ss,t := St − Ss, for (s, t) ∈ ∆[0,T ] := (u, v) ∈ [0, T ]2 : u ≤ v.
PORTFOLIO THEORY WITH ROUGH PATHS 5
For any partition P = 0 = t0 < t1 < · · · < tN = T of an interval [0, T ], we denote the meshsize of P by |P|:= max|tk+1 − tk|: k = 0, 1, . . . , N − 1. A control function is defined as afunction c: ∆[0,T ] → [0,∞) which is superadditive, in the sense that c(s, u) + c(u, t) ≤ c(s, t)
for all 0 ≤ s ≤ u ≤ t ≤ T . For p ∈ [1,∞), the p-variation of a path S ∈ C([0, T ];Rd) over theinterval [s, t] is defined by
‖S‖p,[s,t]:= supP⊂[s,t]
( ∑[u,v]∈P
|Su,v|p)1p
,
where the supremum is taken over all finite partitions P of the interval [s, t], and we usethe abbreviation ‖S‖p:= ‖S‖p,[0,T ]. We say that S has finite p-variation if ‖S‖p< ∞, and
we denote the space of continuous paths with finite p-variation by Cp-var([0, T ];Rd). Notethat S having finite p-variation is equivalent to the existence of a control function c suchthat |Ss,t|p≤ c(s, t) for all (s, t) ∈ ∆[0,T ]. (For instance, one can take c(s, t) = ‖S‖pp,[s,t].)Moreover, for a two-parameter function S: ∆[0,T ] → Rd×d we introduce the correspondingnotion of p-variation by
‖S‖p,[s,t]:= supP⊂[s,t]
( ∑[u,v]∈P
|Su,v|p)1p
,
for p ∈ [1,∞).
Given a k ∈ N and a domain A ⊆ Rd, we will write f ∈ Ck(A), or sometimes simplyf ∈ Ck, to indicate that a function f defined on A is k-times continuously differentiable (inthe Frechet sense), and we will make use of the associated norm
‖f‖Ck := max0≤n≤k
‖Dnf‖∞,
where Dnf denotes the nth order derivative of f , and ‖ · ‖∞ denotes the supremum norm.For a k ∈ N and γ ∈ (0, 1], we will write f ∈ Ck+γ(A), or just f ∈ Ck+γ , to mean that a
function f defined on A is k-times continuously differentiable (in the Frechet sense), and thatits kth order derivative Dkf is locally γ-Holder continuous. In this case we use the norm
‖f‖Ck+γ := max0≤n≤k
‖Dnf‖∞+‖Dkf‖γ-Hol,
where ‖ · ‖γ-Hol denotes the γ-Holder norm.Finally, given two vector spaces U, V , we write L(U ;V ) for the space of linear maps from
U to V .
Let (E, ‖·‖) be a normed space and let f, g:E → R be two functions. We shall write f . gor f ≤ Cg to mean that there exists a constant C > 0 such that f(x) ≤ Cg(x) for all x ∈ E.Note that the value of such a constant may change from line to line, and that the constantsmay depend on the normed space, e.g. through its dimension or regularity parameters.
2.2. Rough path theory and Property (RIE). Let us briefly recall the fundamentaldefinitions of a rough path and of a controlled path, which allow to set up rough integration.
Definition 2.1. For p ∈ (2, 3), a p-rough path is defined as a pair S = (S,S), consisting ofa continuous path S: [0, T ]→ Rd and a continuous two-parameter function S: ∆[0,T ] → Rd×d,such that ‖S‖p<∞, ‖S‖p/2<∞, and Chen’s relation
(2.1) Ss,t = Ss,u + Su,t + Ss,u ⊗ Su,t
6 ALLAN, CUCHIERO, LIU, AND PROMEL
holds for all 0 ≤ s ≤ u ≤ t ≤ T .
Remark 2.2. The success of rough path theory in probability theory is based on the obser-vation that sample paths of many important stochastic processes such as Brownian motion,semimartingales and Markov processes can be enhanced to a rough path, by defining the “en-hancement” S via stochastic integration; see e.g. [FV10, Part III].
Definition 2.3. Let p ∈ (2, 3) and q ≥ p be such that 2/p + 1/q > 1, and let r > 1 be suchthat 1/r = 1/p + 1/q. Let S ∈ Cp-var([0, T ];Rd), F : [0, T ] → Rd and F ′: [0, T ] → L(Rd;Rd)be continuous paths. The pair (F, F ′) is called a controlled path with respect to S (or anS-controlled path), if the Gubinelli derivative F ′ has finite q-variation, and the remainderRF has finite r-variation, where RF : ∆[0,T ] → Rd is defined implicitly by the relation
Fs,t = F ′sSs,t +RFs,t for (s, t) ∈ ∆[0,T ].
We denote the space of controlled paths with respect to S by VqS = VqS([0, T ];Rd), which becomesa Banach space when equipped with the norm
‖F, F ′‖VqS ,[0,T ]:= |F0|+|F ′0|+‖F ′‖q,[0,T ]+‖RF ‖r,[0,T ].
Example 2.4. For a path S ∈ Cp-var([0, T ];Rd) with p ∈ (2, 3), the prototypical example ofa controlled path is (f(S),Df(S)) ∈ VqS for any f ∈ C1+ε with ε ∈ (p − 2, 1] and q = p/ε.Examples of more general controlled paths are discussed in Remark 3.5 and Section 4.1 in thecontext of universal portfolios.
Based on the above definitions, one can establish the existence of the rough integral of acontrolled path (F, F ′) with respect to a p-rough path S. See [FH20] for the correspondingtheory presented in terms of Holder regularity. The following formulation of rough integrationin the language of p-variation can be found in e.g. [PP16, Theorem 4.9].
Theorem 2.5 (Rough integration). Let p ∈ (2, 3) and q ≥ p be such that 2/p+ 1/q > 1, andlet r > 1 be such that 1/r = 1/p+ 1/q. Let S = (S, S) be a p-rough path and let (F, F ′) ∈ VqSbe a controlled path with remainder RF . Then the limit
(2.2)
∫ T
0Fu dSu := lim
|P|→0
∑[s,t]∈P
FsSs,t + F ′sSs,t
exists along every sequence of partitions P of the interval [0, T ] with mesh size |P| tending tozero, and takes values in R. We call this limit the rough integral of (F, F ′) against S. Here,the product FsSs,t is understood as the Euclidean inner product, and the product F ′sSs,t also
takes values in R since the derivative F ′ takes values in L(Rd;Rd) ∼= L(Rd×d;R). Moreover,we have the estimate
(2.3)
∣∣∣∣ ∫ t
sFu dSu − FsSs,t − F ′sSs,t
∣∣∣∣ ≤ C(‖RF ‖r,[s,t]‖S‖p,[s,t]+‖F ′‖q,[s,t]‖S‖ p2,[s,t]),
where the constant C depends only on p, q and r.
In Theorem 2.5 we defined the rough integral of a controlled path (F, F ′) against a roughpath S = (S,S). As noted in [FH20, Remark 4.12], one can actually define a more generalintegral of a controlled path (F, F ′) against another controlled path (G,G′).
PORTFOLIO THEORY WITH ROUGH PATHS 7
Lemma 2.6. Let S = (S, S) be a p-rough path, and let (F, F ′), (G,G′) ∈ VqS be two controlled
paths with remainders RF and RG, respectively. Then the limit
(2.4)
∫ T
0Fu dGu := lim
|P|→0
∑[s,t]∈P
FsGs,t + F ′sG′sSs,t
exists along every sequence of partitions P of the interval [0, T ] with mesh size |P| tending tozero, and comes with the estimate∣∣∣∣ ∫ t
sFu dGu − FsGs,t − F ′sG′sSs,t
∣∣∣∣≤ C
(‖F ′‖∞(‖G′‖qq,[s,t]+‖S‖
pp,[s,t])
1r ‖S‖p,[s,t]+‖F‖p,[s,t]‖RG‖r,[s,t]
+ ‖RF ‖r,[s,t]‖G′‖∞‖S‖p,[s,t]+‖F ′G′‖q,[s,t]‖S‖ p2,[s,t]
),
(2.5)
where the constant C depends only on p, q and r.
Proof. Set Ξs,t := FsGs,t + F ′sG′sSs,t and δΞs,u,t := Ξs,t − Ξs,u − Ξu,t for 0 ≤ s ≤ u ≤ t ≤ T .
Using Chen’s relation (2.1), one can show that
(2.6) δΞs,u,t = −F ′sG′s,uSs,uSu,t − Fs,uRGu,t −RFs,uG′uSu,t − (F ′G′)s,uSu,t.
Since 1/r = 1/p+ 1/q, Young’s inequality gives
|−F ′sG′s,uSs,uSu,t| ≤ ‖F ′‖∞‖G′‖q,[s,u]‖S‖p,[s,u]‖S‖p,[u,t]
. ‖F ′‖∞(‖G′‖qq,[s,u]+‖S‖pp,[s,u])
1r ‖S‖p,[u,t]= w1(s, u)
1rw2(u, t)
1p ,
where w1(s, u) := ‖F ′‖r∞(‖G′‖qq,[s,u]+‖S‖pp,[s,u]) and w2(u, t) := ‖S‖pp,[u,t] are control func-
tions. Treating the other three terms on the right-hand side of (2.6) similarly, we deducethe hypotheses of the generalized sewing lemma [FZ18, Theorem 2.5], from which the resultfollows.
Rough integration offers strong pathwise stability estimates, and may be viewed as arguablythe most general pathwise integration theory, generalizing classical notions of integration suchas those of Riemann–Stieltjes, Young and Follmer, and allowing one to treat many well-known stochastic processes as integrators; see e.g. [FH20]. However, from the perspectiveof mathematical finance, rough integration comes with one apparent flaw: the definition ofrough integral (2.2) is based on so-called “compensated” Riemann sums, and thus does not(at first glance) come with the natural interpretation as the capital gain process associated toan investment in a financial market. Indeed, let us suppose that S represents the asset priceson a financial market and F an investment strategy. In this case, neither the associated roughpath S = (S, S) nor the controlled path (F, F ′), assuming they exist, are uniquely determined
by S and F , but rather the value of the rough integral∫ T
0 Fu dSu will depend in general onthe choices of S and F ′. Moreover, the financial meaning of the term F ′sSs,t appearing in thecompensated Riemann sum in (2.2) is far from obvious.
As observed in [PP16], the aforementioned drawback of rough integration from a financialperspective can be resolved by introducing the following property of the price path S.
8 ALLAN, CUCHIERO, LIU, AND PROMEL
Property (RIE). Let p ∈ (2, 3) and let Pn = 0 = tn0 < tn1 < · · · < tnNn = T, n ∈ N,be a sequence of partitions of the interval [0, T ], such that |Pn|→ 0 as n → ∞. For S ∈C([0, T ];Rd), we define Sn: [0, T ]→ Rd by
Snt := ST1T(t) +
Nn−1∑k=0
Stnk1[tnk ,tnk+1)(t), t ∈ [0, T ],
for each n ∈ N. We assume that:
• the Riemann sums∫ t
0 Snu ⊗ dSu :=
∑Nn−1k=0 Stnk ⊗ Stnk∧t,t
nk+1∧t converge uniformly as
n→∞ to a limit, which we denote by∫ t
0 Su ⊗ dSu, t ∈ [0, T ],
• and that there exists a control function c such that1
sup(s,t)∈∆[0,T ]
|Ss,t|p
c(s, t)+ supn∈N
sup0≤k<`≤Nn
|∫ tn`tnkSnu ⊗ dSu − Stnk ⊗ Stnk ,tn` |
p2
c(tnk , tn` )
≤ 1.
Definition 2.7. A path S ∈ C([0, T ];Rd) is said to satisfy (RIE) with respect to p and(Pn)n∈N, if p, (Pn)n∈N and S together satisfy Property (RIE).
As discussed in detail in [PP16], if a path S ∈ C([0, T ];Rd) satisfies (RIE) with respect top and (Pn)n∈N, then S can be enhanced to a p-rough path S = (S,S) by setting
(2.7) Ss,t :=
∫ t
sSu ⊗ dSu − Ss ⊗ Ss,t, for (s, t) ∈ ∆[0,T ].
In other words, Property (RIE) ensures the existence of a rough path associated to thepath S. The advantage of the (more restrictive) Property (RIE) is that it guarantees thatthe corresponding rough integrals can be well approximated by classical left-point Riemannsums, as we will see in Section 2.4, thus allowing us to restore the financial interpretation ofsuch integrals as capital processes.
Remark 2.8. The assumption that the underlying price paths satisfy Property (RIE) appearsto be rather natural in the context of portfolio theory. Indeed, in stochastic portfolio theorythe price processes are commonly modelled as semimartingales fulfilling the condition of “nounbounded profit with bounded risk” (NUPBR); see e.g. [Fer02]. The condition (NUPBR) isalso essentially the minimal condition required to ensure that expected utility maximizationproblems are well-posed; see [KK07, IP15]. As established in [PP16, Proposition 2.7 andRemark 4.16], the sample paths of semimartingales fulfilling (NUPBR) almost surely satisfyProperty (RIE) with respect to every p ∈ (2, 3) and a suitably chosen sequence of partitions.
2.3. The bracket process and a rough Ito formula. A vital tool in many applicationsof stochastic calculus is Ito’s formula, and it will also be an important ingredient in ourcontribution to portfolio theory. Usually, (pathwise) Ito formulae are based on the notion ofquadratic variation. In rough path theory, a similar role as that of the quadratic variation isplayed by the so-called bracket of a rough path, cf. [FH20, Definition 5.5].
Definition 2.9. Let S = (S, S) be a p-rough path and let Sym(S) denote the symmetric partof S. The bracket of S is defined as the path [S]: [0, T ]→ Rd×d given by
[S]t := S0,t ⊗ S0,t − 2Sym(S0,t), t ∈ [0, T ].
1Here and throughout, we adopt the convention that 00
:= 0.
PORTFOLIO THEORY WITH ROUGH PATHS 9
The bracket of a rough path allows one to derive Ito formulae for rough paths. For thispurpose, note that [S] is a continuous path of finite p/2-variation, which can be seen from theobservation that
[S]s,t = [S]t − [S]s = Ss,t ⊗ Ss,t − 2Sym(Ss,t), for all (s, t) ∈ ∆[0,T ].
The following Ito formula for rough paths can be proven almost exactly as the one in [FH20,Theorem 7.7], so we will omit its proof here; see also [FZ18, Theorem 2.12].
Proposition 2.10. Let S = (S, S) be a p-rough path and let Γ ∈ Cp2
-var([0, T ];Rd). Supposethat F, F ′ and F ′′ are such that (F, F ′), (F ′, F ′′) ∈ VqS, and F =
∫ ·0 F′u dSu + Γ. If g ∈ Cp+ε
for some ε > 0, then, for every t ∈ [0, T ], we have
g(Ft) = g(F0) +
∫ t
0Dg(Fu)F ′u dSu +
∫ t
0Dg(Fu) dΓu +
1
2
∫ t
0D2g(Fu)(F ′u ⊗ F ′u) d[S]u.
Assuming Property (RIE), it turns out that the bracket [S] of a rough path S = (S,S) doescoincide precisely with the quadratic variation of the path S in the sense of Follmer [Fol81].
Lemma 2.11. Suppose that S ∈ C([0, T ];Rd) satisfies (RIE) with respect to p and (Pn)n∈N.Let S = (S,S) be the associated rough path as defined in (2.7). Then, the bracket [S] has finitetotal variation, and is given by
[S]t = limn→∞
Nn−1∑k=0
Stnk∧t,tnk+1∧t ⊗ Stnk∧t,tnk+1∧t,
where the convergence is uniform in t ∈ [0, T ].
Proof. The (i, j)-component of [S]t is given by
[S]ijt = Si0,tSj0,t − Sij0,t − Sji0,t = SitS
jt − Si0S
j0 −
∫ t
0Siu dSju −
∫ t
0Sju dSiu.
The result then follows from Lemmas 4.17 and 4.22 in [PP16].
In view of Lemma 2.11, when assuming Property (RIE), we also refer to the bracket [S] asthe quadratic variation of S.
2.4. Rough integrals as limits of Riemann sums. As previously mentioned, the mainmotivation to introduce Property (RIE) is to obtain the rough integral as a limit of left-pointRiemann sums, in order to restore the interpretation of the rough integral as the capitalprocess associated with a financial investment. Indeed, we present the following extension of[PP16, Theorem 4.19], which will be another central tool in our pathwise portfolio theory.The proof of Theorem 2.12 is postponed to Appendix B.
Theorem 2.12. Suppose that S ∈ C([0, T ];Rd) satisfies (RIE) with respect to p and (Pn)n∈N.Let q ≥ p such that 2/p + 1/q > 1. Let f ∈ Cp+ε for some ε > 0, so that in particular(f(S),Df(S)) ∈ VqS. Then, for any (Y, Y ′) ∈ VqS, the integral of (Y, Y ′) against (f(S),Df(S)),as defined in Lemma 2.6, is given by
(2.8)
∫ t
0Yu df(S)u = lim
n→∞
Nn−1∑k=0
Ytnk f(S)tnk∧t,tnk+1∧t,
where the convergence is uniform in t ∈ [0, T ].
10 ALLAN, CUCHIERO, LIU, AND PROMEL
As an immediate consequence of Theorem 2.12, assuming Property (RIE), we note that,for (Y, Y ′) ∈ VqS , the rough integral
(2.9)
∫ t
0Yu dSu = lim
n→∞
Nn−1∑k=0
YtnkStnk∧t,t
nk+1∧t,
and indeed the more general rough integral in (2.8), is independent of the Gubinelli derivativeY ′. However, in the spirit of Follmer’s pathwise quadratic variation and integration, the right-hand sides of (2.8) and (2.9) do in general depend on the sequence of partitions (Pn)n∈N.
3. Pathwise (relative) portfolio wealth processes and master formula
In this section we consider pathwise portfolio theory on the rough path foundation presentedin Section 2. In particular, we study the growth of wealth processes relative to the marketportfolio, and provide an associated pathwise master formula analogous to that of classicalstochastic portfolio theory, cf. [Fer99, Str14, SSV18]. We start by introducing the basicassumptions on the underlying financial market.
3.1. The financial market. Since we want to investigate the long-run behaviour of wealthprocesses, we consider the price trajectories of d assets on the time interval [0,∞). As iscommon in stochastic portfolio theory, we do not include default risk—that is, all prices areassumed to be strictly positive—and we do not distinguish between risk-free and risky assets.
A partition P of the interval [0,∞) is a strictly increasing sequence of points (ti)i≥0 ⊂[0,∞), with t0 = 0 and such that ti →∞ as i→∞. Given any T > 0, we denote by P([0, T ])the restriction of the partition P∪T to the interval [0, T ], i.e. P([0, T ]) := (P∪T)∩ [0, T ].For a path S: [0,∞) → Rd, we write S|[0,T ] for the restriction of S to [0, T ], and we setR+ := (0,∞).
Definition 3.1. For a fixed p ∈ (2, 3), we say that a path S ∈ C([0,∞);Rd+) is a price path,if there exists a sequence of partitions (PnS )n∈N of the interval [0,∞), with vanishing meshsize on compacts, such that, for all T > 0, the restriction S|[0,T ] satisfies (RIE) with respectto p and (PnS ([0, T ]))n∈N.
We denote the family of all such price paths by Ωp.
It seems to be natural to allow the partitions (PnS )n∈N to depend on the price path S, sincepartitions are typically given via stopping times in stochastic frameworks.
Throughout the remainder of the paper, we adopt the following assumption on the regu-larity parameters.
Assumption 3.2. Let p ∈ (2, 3), q ≥ p and r > 1 be given such that
2
p+
1
q> 1 and
1
r=
1
p+
1
q.
In particular, we note that 1 < p/2 ≤ r < p ≤ q <∞.
By Property (RIE), we can (and do) associate to every price path S ∈ Ωp the p-roughpath S = (S,S), as defined in (2.7). We can then define the market covariance as the matrixa = [aij ]1≤i,j≤d, with (i, j)-component given by the measure
(3.1) aij(ds) :=1
SisSjs
d[S]ijs .
PORTFOLIO THEORY WITH ROUGH PATHS 11
Although we do not work in a probabilistic setting and thus should not, strictly speaking,talk about covariance in the probabilistic sense, the relation (3.1) is consistent with classicalstochastic portfolio theory (with the bracket process replaced by the quadratic variation), andit turns out to still be a useful quantity in pathwise frameworks, cf. [SV16, SSV18].
3.2. Pathwise portfolio wealth processes. We now introduce admissible portfolios andthe corresponding wealth processes on the market defined above. To this end, we first fix thenotation:
∆d :=
x = (x1, . . . , xd) ∈ Rd :
d∑i=1
xi = 1
,
∆d+ := x ∈ ∆d : xi > 0 ∀i = 1, . . . , d and ∆
d+ := x ∈ ∆d : xi ≥ 0 ∀i = 1, . . . , d.
Definition 3.3. We say that a path F : [0,∞) → Rd is an admissible strategy if, for everyT > 0, there exists a path F ′: [0, T ] → L(Rd;Rd) such that (F |[0,T ], F
′) ∈ VqS is a controlledpath with respect to S (in the sense of Definition 2.3). We say that an admissible strategy πis a portfolio for S if additionally πt ∈ ∆d for all t ∈ [0,∞).
Remark 3.4. As explained in [FH20, Remark 4.7], if S is sufficiently regular then, given anadmissible strategy F , there could exist multiple different Gubinelli derivatives F ′ such thatthe pair (F, F ′) defines a valid controlled path with respect to S. However, thanks to Property(RIE), Theorem 2.12 shows that the rough integral
∫F dS can be expressed as a limit of
Riemann sums which only involve F and S, and, therefore, is independent of the choice ofF ′. Thus, the choice of the Gubinelli derivative F ′ is unimportant, provided that at least oneexists. Indeed, one could define an equivalence relation ∼ on VqS such that (F, F ′) ∼ (G,G′) ifF = G, and define the family of admissible strategies as elements of the quotient space VqS/∼.By a slight abuse of notation, we shall therefore sometimes write simply F ∈ VqS instead of(F, F ′) ∈ VqS.
Remark 3.5. While the admissible class of portfolios introduced in Definition 3.3 allows fora pathwise (model-free) analysis (without notions like filtration or predictability), it also cov-ers the most frequently applied classes of functionally generated portfolios—see [Fer99]—andtheir generalizations as considered in e.g. [Str14] and [SSV18]. Indeed, every path-dependentfunctionally generated portfolio which is sufficiently smooth in the sense of Dupire [Dup19](see also [CF10]), is a controlled path and thus an admissible strategy, as shown in [Ana20].
In the present work we will principally focus on “adapted” strategies F , in the sense thatF is a controlled path, as in Definition 3.3, with Ft being a measurable function of S|[0,t] foreach t ∈ [0,∞). In other words, if S is modelled by a stochastic process then we require F tobe adapted to the natural filtration generated by S. Clearly, such adapted admissible strategiesare reasonable choices in the context of mathematical finance.
A portfolio π = (π1, . . . , πd) represents the ratio of the investor’s wealth invested into eachof the d assets. As is usual, we normalize the initial wealth to be 1, since in the followingwe will only be concerned with the long-run growth. Suppose S ∈ Ωp with correspondingsequence (PnS )n∈N of partitions. If we restrict the rebalancing according to the portfolio π tothe discrete times given by PnS = (tnj )j∈N, then the corresponding wealth process Wn satisfies
Wnt = 1 +
∞∑j=1
πtjWntj
StjStj∧t,tj+1∧t = 1 +
∞∑j=1
d∑i=1
πitjWntj
SitjSitj∧t,tj+1∧t
12 ALLAN, CUCHIERO, LIU, AND PROMEL
with tj ∧ t := mintj , t. Taking the limit to continuous-time (i.e. n → ∞) and keepingProperty (RIE) in mind, we observe that the wealth process W π associated to the portfolioπ should satisfy
(3.2) W πt = 1 +
∫ t
0
πsWπs
SsdSs, t ∈ [0,∞).
Analogously to (classical) stochastic portfolio theory (e.g. [KK07] or [SSV18]), the wealthprocess associated to a portfolio may be expressed as a (rough) exponential.
Lemma 3.6. Let π be a portfolio for S ∈ Ωp. Then the wealth process W π (with unit initialwealth), given by
W πt := exp
(∫ t
0
πsSs
dSs −1
2
d∑i,j=1
∫ t
0
πisπjs
SisSjs
d[S]ijs
), t ∈ [0,∞),
satisfies (3.2), where∫ t
0πsSs
dSs is the rough integral of the controlled path π/S with respect
to rough path S, and∫ t
0πisπ
js
SisSjs
d[S]ijs is the usual Riemann–Stieltjes integral with respect to the
(i, j)-component of the (finite variation) bracket [S].
Proof. Note that, since 1/S = f(S) with the smooth function f(x) = (1/x1, . . . , 1/xd) on Rd+,the pair (1/S,Df(S)) ∈ VpS ⊂ V
qS is a controlled path. Therefore, for each portfolio π ∈ VqS ,
we can define the quotient π/S = (π1/S1, . . . , πd/Sd), which gives an element (π/S, (π/S)′)in VqS ; see Lemma A.1.
Setting Z :=∫ ·
0πsSs
dSs, by Lemma B.1, we have that
[Z] =
∫ ·0
(πsSs⊗ πsSs
)d[S]s =
d∑i,j=1
∫ ·0
πisπjs
SisSjs
d[S]ijs ,
where Z is the canonical rough path lift of Z (see Section A.3). We then have that W πt =
exp(Zt − 12 [Z]t), so that, by Lemma A.5, W π satisfies
W πt = 1 +
∫ t
0W πs dZs, t ∈ [0,∞).
By Lemma A.4 and Proposition A.2 it then follows that W π satisfies (3.2).
Remark 3.7. Every portfolio π can be associated to a self-financing admissible strategy ξ by
setting ξit := πitWπt /S
it for i = 1, . . . , d. Indeed, we have that W π
t =∑d
i=1 ξitS
it, and that
W πt = 1 +
∫ t
0
πsWπs
SsdSs = 1 +
∫ t
0ξs dSs, t ∈ [0,∞),
so that ξ is self-financing.
As in the classical setup of stochastic portfolio theory (e.g. [Fer02]) we introduce the marketportfolio as a reference portfolio.
Lemma 3.8. The path µ: [0,∞)→ ∆d+, defined by µit :=
SitS1t+···+Sdt
for i = 1, . . . , d, is a port-
folio for S ∈ Ωp, called the market portfolio (or market weights process). The corresponding
PORTFOLIO THEORY WITH ROUGH PATHS 13
wealth process (with initial wealth 1) is given by
Wµt =
S1t + · · ·+ SdtS1
0 + · · ·+ Sd0.
Proof. Since µ is a smooth function of S, it is a controlled path with respect to S, and istherefore an admissible strategy. Since µ1
t + · · ·+ µdt = 1, we see that µ is indeed a portfolio.Let f(x) := log(x1 + · · · + xd) for x ∈ Rd+. By the Ito formula for rough paths (Proposi-
tion 2.10), it follows that
f(St)− f(S0) =
∫ t
0
(1
S1s + · · ·+ Sds
, . . . ,1
S1s + · · ·+ Sds
)dSs −
1
2
∫ t
0
(µsSs⊗ µsSs
)d[S]s
=
∫ t
0
µsSs
dSs −1
2
d∑i,j=1
∫ t
0
µisµjs
SisSjs
d[S]ijs ,
where we used the fact that µisSis
= 1S1s+···+Sds
. By Lemma 3.6, the right-hand side is equal to
logWµt , so that
Wµt = exp (f(St)− f(S0)) =
S1t + · · ·+ SdtS1
0 + · · ·+ Sd0.
3.3. Formulae for the growth of wealth processes. In this subsection we derive pathwiseversions of classical formulae of stochastic portfolio theory—see [Fer99]—which describe thedynamics of the relative wealth of a portfolio with respect to the market portfolio; cf. [SSV18]for analogous results relying on Follmer’s pathwise integration.
Given a portfolio π, we define the relative covariance of π by τπ = [τπij ]1≤i,j≤d, where
(3.3) τπij(ds) := (πs − ei)>a(ds)(πs − ej),
where (ei)1≤i≤d denotes the canonical basis of Rd, and we recall a(ds) as defined in (3.1).Henceforth, we will write
(3.4) V π :=W π
Wµ
for the relative wealth of a portfolio π with respect to the market portfolio µ.
Proposition 3.9. Let π be a portfolio for S ∈ Ωp, and let µ be the market portfolio as above.We then have that
(3.5) log V πt =
∫ t
0
πsµs
dµs −1
2
d∑i,j=1
∫ t
0πisπ
jsτµij(ds), t ∈ [0,∞).
Remark 3.10. The integral∫ t
0πsµs
dµs appearing in (3.5) is interpreted as the rough integral
of the S-controlled path π/µ against the S-controlled path µ in the sense of Lemma 2.6. By
Theorem 2.12, the integral∫ t
0πsµs
dµs can also be expressed as a limit of left-point Riemann
sums, which justifies the financial meaning of (3.5).
14 ALLAN, CUCHIERO, LIU, AND PROMEL
Proof of Proposition 3.9. Step 1. By the Ito formula for rough paths (Proposition 2.10), with
the usual notational convention log x =∑d
i=1 log xi, we have
logSt = logS0 +
∫ t
0
1
SsdSs −
1
2
d∑i=1
∫ t
0
1
(Sis)2
d[S]iis , t ∈ [0,∞).
Since π and logS are S-controlled paths, we can define the integral of π against logS in thesense of Lemma 2.6. By the associativity of rough integration (Proposition A.2), we have∫ t
0πs d logSs =
∫ t
0
πsSs
dSs −1
2
d∑i=1
∫ t
0
πis(Sis)
2d[S]iis .
It is convenient to introduce the excess growth rate of the portfolio π, given by
γ∗π(ds) :=1
2
( d∑i=1
πisaii(ds)−
d∑i,j=1
πisπjsaij(ds)
).
By Lemma 3.6, we have that
(3.6) logW πt =
∫ t
0
πsSs
dSs −1
2
d∑i,j=1
∫ t
0πisπ
jsaij(ds) =
∫ t
0πs d logSs + γ∗π([0, t]).
In particular, this implies that
(3.7) log V πt =
∫ t
0(πs − µs) d logSs + γ∗π([0, t])− γ∗µ([0, t]).
Step 2. By Lemma 3.8 and (3.6), we have
logµit = logµi0 + logSit − logSi0 − logWµt
= logµi0 + logSit − logSi0 −∫ t
0µs d logSs − γ∗µ([0, t])
= logµi0 +
∫ t
0(ei − µs) d logSs − γ∗µ([0, t]).(3.8)
By part (ii) of Proposition B.2 and Lemma B.1, we deduce that
(3.9) [logS]t = a([0, t]), and [logµ]t = τµ([0, t]).
Applying the Ito formula for rough paths (Proposition 2.10) to exp(logµi), using the associa-tivity of rough integration (Proposition A.2), and recalling (3.8), we have∫ t
0
πisµis
dµis =
∫ t
0πis(ei − µs) d logSs −
∫ t
0πis dγ∗µ(ds) +
1
2
∫ t
0πis d[logµ]iis .
Using (3.9) and summing over i = 1, . . . , d, we obtain
(3.10)
∫ t
0
πsµs
dµs =
∫ t
0(πs − µs) d logSs − γ∗µ([0, t]) +
1
2
d∑i=1
∫ t
0πisτ
µii(ds).
Step 3. Taking the difference of (3.7) and (3.10), we have
log V πt =
∫ t
0
πsµs
dµs + γ∗π([0, t])− 1
2
d∑i=1
∫ t
0πisτ
µii(ds).
PORTFOLIO THEORY WITH ROUGH PATHS 15
It remains to note that
γ∗π([0, t]) =1
2
( d∑i=1
∫ t
0πisτ
µii(ds)−
d∑i,j=1
∫ t
0πisπ
jsτµij(ds)
),
which follows from a straightforward calculation; see e.g. [Fer02, Lemma 1.3.4].
While Definition 3.3 allows for rather general portfolios, so-called functionally generatedportfolios are the most frequently considered ones in SPT. In a pathwise setting such portfoliosand the corresponding master formula were studied previously in [SSV18] and [CSW19]. Weconclude this section by deriving such a master formula for functionally generated portfoliosin the present (rough) pathwise setting.
Let G be a strictly positive function in Cp+ε(∆d+;R+) for some ε > 0. One can verify that
∇ logG(µ) ∈ Vqµ is a µ-controlled path for a suitable choice of q (see Example 2.4), and istherefore also an S-controlled path by Lemma A.4. Since the product of controlled paths isitself a controlled path (by Lemma A.1), we see that the path π defined by
(3.11) πit := µit
(∂
∂xilogG(µt) + 1−
d∑k=1
µkt∂
∂xklogG(µt)
), t ∈ [0,∞), i = 1, . . . , d,
is a µ-controlled (and hence also an S-controlled) path, and is indeed a portfolio for S ∈ Ωp.The function G is called a portfolio generating function, and we say that G generates π.
Theorem 3.11 (The master formula). Let G ∈ Cp+ε(∆d+;R+) for some ε > 0 be a portfolio
generating function, and let π be the portfolio generated by G. The wealth of π relative to themarket portfolio is given by
log V πt = log
(G(µt)
G(µ0)
)− 1
2
d∑i,j=1
∫ t
0
1
G(µs)
∂2G(µs)
∂xi∂xjµisµ
jsτµij(ds), t ∈ [0,∞).
Proof. Let g = ∇ logG(µ), so that gi = ∂∂xi
logG(µ) = 1G(µ)
∂G∂xi
(µ) for each i = 1, . . . , d. We
can then rewrite (3.11) as
(3.12) πi = µi(gi + 1−
d∑k=1
µkgk),
so that πi/µi = gi + 1 −∑d
k=1 µkgk. Since
∑di=1 µ
is = 1 for all s ≥ 0, we must have that∑d
i=1 µis,t = 0 for all s < t. Thus∫ t
0
πsµs
dµs = limn→∞
Nn−1∑k=0
d∑i=1
πitnkµitnk
µitnk∧t,tnk+1∧t
= limn→∞
Nn−1∑k=0
d∑i=1
gitnkµitnk∧t,t
nk+1∧t
=
∫ t
0gs dµs.
We have from (3.3) that∑d
j=1 µjsτµij(ds) = (µs − ei)>a(ds)(µs − µs) = 0. It follows from this
and (3.12) that
(3.13)
d∑i,j=1
πisπjsτµij(ds) =
d∑i,j=1
gisgjsµ
isµjsτµij(ds).
16 ALLAN, CUCHIERO, LIU, AND PROMEL
Recall from (3.9) that [logµ]t = τµ([0, t]). By applying the Ito formula for rough paths
(Proposition 2.10) to µi = exp(logµi), we see that the path t 7→ µit −∫ t
0 µis d logµis is of finite
variation. By part (ii) of Proposition B.2 and Lemma B.1, we therefore have that
(3.14) [µ]ijt =
∫ t
0µisµ
js d[logµ]ijs =
∫ t
0µisµ
jsτµij(ds).
By the Ito formula for rough paths (Proposition 2.10), we then have
log
(G(µt)
G(µ0)
)=
∫ t
0gs dµs +
1
2
d∑i,j=1
∫ t
0
(1
G(µs)
∂2G(µs)
∂xi∂xj− gisgjs
)d[µ]ijs
=
∫ t
0
πsµs
dµs +1
2
d∑i,j=1
∫ t
0
(1
G(µs)
∂2G(µs)
∂xi∂xj− gisgjs
)µisµ
jsτµij(ds).
Combining this with (3.5) and (3.13), we deduce the result.
4. Cover’s universal portfolios and their optimality
Like stochastic portfolio theory, Cover’s universal portfolios [Cov91] aim to give generalrecipes to construct preference-free asymptotically “optimal” portfolios; see also [Jam92] and[CO96]. A first link between SPT and these universal portfolios was established in a pathwiseframework based on Follmer integration in [CSW19] (see also [Won15]). In this sectionwe shall generalize the pathwise theory regarding Cover’s universal portfolios developed in[CSW19] to the present rough path setting.
Cover’s universal portfolio is based on the idea of trading according to a portfolio which isdefined as the average over a family A of admissible portfolios. In the spirit of [CSW19], weintroduce pathwise versions of Cover’s universal portfolios—that is, portfolios of the form
πνt :=
∫A πtV
πt dν(π)∫
A Vπt dν(π)
, t ∈ [0,∞),
where ν is a given probability measure on A. In order to find suitable classes A of admissibleportfolios, we recall Assumption 3.2 and make the following standing assumption throughoutthe entire section.
Assumption 4.1. We fix q′ > q and r′ > r such that 2p + 1
q′ > 1 and 1r′ = 1
p + 1q′ .
4.1. Admissible portfolios. As a first step to construct Cover’s universal portfolios in ourrough path setting, we need to find a suitable set of admissible portfolios. To this end, we set
Vqµ([0,∞); ∆d) :=
(π, π′) : ∀T > 0, (π, π′)|[0,T ]∈ Vqµ([0, T ]; ∆d).
Then, for some fixed control function cµ which controls the p-variation norm of the marketportfolio µ, and for some M > 0, we introduce a class of admissible portfolios as the set
(4.1) AM,q(cµ) :=
(π, π′) ∈ Vqµ([0,∞); ∆d) :
∣∣∣π0µ0 ∣∣∣+∣∣∣(πµ)′0∣∣∣ ≤M,
sups≤t|(πµ
)′s,t|q
cµ(s,t) + sups≤t|R
πµs,t|r
cµ(s,t) ≤ 1
.
Here (π/µ, (π/µ)′) denotes the product of the two µ-controlled paths (π, π′) and ( 1µ , (
1µ)′)
(see Lemma A.1). In particular, (π/µ)′ = π′/µ + π(1/µ)′, and Rπµ is the remainder of the
controlled rough path π/µ.
PORTFOLIO THEORY WITH ROUGH PATHS 17
Remark 4.2. We consider here controlled paths with respect to µ, instead of with respect toS. As noted in Remark 3.10, every S-controlled path (π, π′) ∈ VqS can be used to define theintegral
∫πtµt
dµt, and all the results in this section can also be established based on VqS with
appropriate modifications. We choose to consider (π, π′) ∈ Vqµ as a µ-controlled path in orderto slightly simplify the notation. It is straightforward to check that Vqµ ⊆ VqS.
Let us recall from Definition 2.3 that, for any T > 0,
‖(Y, Y ′)‖Vqµ,[0,T ]= |Y0|+|Y ′0 |+‖Y ′‖q,[0,T ]+‖RY ‖r,[0,T ]
defines a complete norm on Vqµ([0, T ]; ∆d). We endow AM,q(cµ) ⊂ Vq′µ ([0,∞); ∆d) with the
seminorms
(4.2) pµ,q′
T ((π, π′)) :=∥∥∥πµ,(πµ
)′∥∥∥Vq′µ ,[0,T ]
, T > 0.
The reason for taking q′ > q is that it will allow us to obtain a compact embedding of AM,q(cµ)
into Vq′µ . This compactness of the set of admissible portfolios plays a crucial role in obtaining
optimality of universal portfolios.
Let us discuss some examples of admissible portfolios. We first check that the function-ally generated portfolios treated in [CSW19] belong to AM,q(cµ) provided that the control
function cµ is chosen appropriately. Recall that Ck(∆d+;R+) denotes the space of k-times con-
tinuously differentiable R+-valued functions on the closed (non-negative) simplex ∆d+, and
that ‖G‖Ck := max0≤n≤k‖DnG‖∞.
Lemma 4.3. Let K > 0 be a constant, and let
GK =G ∈ C3(∆
d+;R+) : ‖G‖C3≤ K, G ≥
1
K
.
Then the portfolio π generated by G, as defined in (3.11), belongs to AM,p(cµ) for a suitablecontrol function cµ and constant M . More precisely, there exists a control function of theform cµ( · , ·) = C‖µ‖pp,[ · , · ] and a constant M > 0, such that C and M only depend on K, and
(πG, (πG)′) : πG defined in (3.11) for some G ∈ GK⊂ AM,p(cµ).
Note that here we take q = p and r = p/2.
Proof. Fix G ∈ GK , and let π be the associated portfolio as defined in (3.11). Since π isdefined as a C2 function of µ, we know immediately that it is a µ-controlled path.
A simple calculation shows thatπtµt
= gt + (1− µt · gt)1,
where we write 1 = (1, . . . , 1) and gt = ∇ logG(µt), and we use · to denote the standardinner product on Rd. The pair (1, 0) is trivially a µ-controlled path with 1′ = 0 and R1 = 0,and thus clearly satisfies the required bounds in (4.1) with an arbitrary control function. Itthus suffices to show that (g, g′) and (µ · g, (µ · g)′) satisfy the required bounds with controlfunctions c1
µ and c2µ respectively, since then cµ := c1
µ + c2µ gives the desired control function.
We begin with (g, g′). Let F := ∇ logG, so that g = F (µ) and g′ = DF (µ). By Taylorexpansion, we can verify that, for all s ≤ t,(4.3) |gs,t|≤ ‖DF‖∞|µs,t|, |g′s,t|≤ ‖D2F‖∞|µs,t|, |Rgs,t|≤ ‖D2F‖∞|µs,t|2.
18 ALLAN, CUCHIERO, LIU, AND PROMEL
Note that F , DF and D2F only depend on DG, D2G, D3G and 1/G, and therefore, since‖G‖C3≤ K and G ≥ 1/K, there exists a constant C = C(K), which only depends on K,such that ‖F‖∞≤ C, ‖DF‖∞≤ C and ‖D2F‖∞≤ C. It follows that we can choose c1
µ(s, t) =
C‖µ‖pp,[s,t]. Note also that ‖g‖∞≤ C and ‖g′‖∞≤ C.
We now turn to (µ · g, (µ · g)′). Noting that µ is trivially a µ-controlled path with µ′ = 1and Rµ = 0, and that Rµ·gs,t = µs ·Rgs,t + µs,t · gs,t, we deduce that
|(µ · g)′s,t|≤ |gs,t|+‖µ‖∞|g′s,t|+‖g′‖∞|µs,t|, |Rµ·gs,t |≤ ‖µ‖∞|Rgs,t|+|µs,t||gs,t|.
Since µt takes values in the bounded set ∆d+, we can use the bounds in (4.3) to show that
there exists a constant L = L(K), depending only on K, such that |(µ · g)′s,t|≤ L|µs,t| and
|Rµ·gs,t |≤ L|µs,t|2. It follows that we may take c2µ(s, t) := L‖µ‖pp,[s,t]. Finally, we note that
the initial values π0/µ0 = g0 + (1 − µ0 · g0)1 = F (µ0) + (1 − µ0 · F (µ0))1 and (π/µ)′0 =DF (µ0)− (F (µ0)+µ0DF (µ0))1 are also bounded by a constant M depending only on K.
One particular advantage of rough integration is that the admissible strategies need notbe of gradient type, giving us more flexibility in choosing admissible portfolios compared toprevious approaches relying on Follmer integration.
Example 4.4 (Functionally controlled portfolios). Let
F2,K :=
(πF , πF,′) : F ∈ C2(∆d+;Rd), ‖F‖C2≤ K
for a given constant K > 0, where
(πFt )i = µit
(F i(µt) + 1−
d∑j=1
µjtFj(µt)
)(4.4)
for t ≥ 0 and i = 1, . . . , d. Then F2,K ⊂ AM,p(cµ), where we can again take q = p. Thepoint here is that we can consider all C2-functions F , rather than requiring that F is of theform F = ∇ logG for some function G. One can verify that F2,K ⊂ AM,p(cµ) for a suitablecontrol function cµ by following the proof of Lemma 4.3 almost verbatim.
Example 4.5 (Controlled equation generated portfolios). Let us define
C3,K := f ∈ C3(Rd;L(Rd;Rd)) : ‖f‖C3≤ K.
For a given f ∈ C3,K , a classical result in rough path theory is that the controlled differentialequation with the vector field f , driven by µ,
(4.5) dY ft = f(Y f
t ) dµt, Y0 = ξ ∈ ∆d,
admits a unique solution (Y f , (Y f )′) = (ξ+∫ ·
0 f(Y fu ) dµu, f(Y f )), which is itself a µ-controlled
path. Moreover, writing Aµs,t =∫ ts µs,u ⊗ dµu for the canonical rough path lift of µ (see
Section A.3), and cµ(s, t) := ‖µ‖pp,[s,t]+‖Aµ‖
p2p2,[s,t]
, for every T > 0, there exists a constant ΓT
depending on p, cµ([0, T ]) and K, such that
sup(s,t)∈∆[0,T ]
|(Y f )′s,t|p
ΓT cµ(s, t)+ sup
(s,t)∈∆[0,T ]
|RY fs,t |p2
ΓT cµ(s, t)≤ 1.
PORTFOLIO THEORY WITH ROUGH PATHS 19
Consequently, as in the proof of Lemma 4.3 one can show that there exists an increasingfunction Γ: [0,∞)→ R+, depending on p, cµ and K such that
sup0≤s≤t<∞
|(πfµ )′s,t|p
cµ(s, t)+ sup
0≤s≤t<∞
|Rπf
µ
s,t |p2
cµ(s, t)≤ 1,
where πf := µ(Y f + (1− µ · Y f )1) and cµ(s, t) := Γtcµ(s, t) is again a control function. Thisimplies that the set
πf = µ(Y f + (1− µ · Y f )1) : Y f is the solution of (4.5) for some f ∈ C3,K ⊂ AM,p(cµ)
for a suitable constant M > 0.
4.2. Asymptotic growth of universal portfolios. To investigate the asymptotic growthrates of our pathwise versions of Cover’s universal portfolio, we first require some auxiliaryresults—in particular the compactness of the set of admissible portfolios.
Lemma 4.6. The set AM,q(cµ) is compact in the topology generated by the family of semi-
norms pµ,q′
T : T ∈ N as defined in (4.2), where we recall that q < q′.
Proof. Step 1 : We first show that the set
A :=
(Y, Y ′) ∈ Vqµ([0,∞);Rd) : |Y0|+|Y ′0 |≤M and sup
s≤t
|Y ′s,t|q
cµ(s, t)+ sup
s≤t
|RYs,t|r
cµ(s, t)≤ 1
is compact with respect to the topology generated by the seminorms ‖· , ·‖Vq′µ ,[0,T ]
for T ∈ N.
It suffices to show that for every fixed T ∈ N, the set
AT :=
(Y, Y ′) ∈ Vqµ([0, T ];Rd) : |Y0|+|Y ′0 |≤M and
sup(s,t)∈∆[0,T ]
|Y ′s,t|q
cµ(s, t)+ sup
(s,t)∈∆[0,T ]
|RYs,t|r
cµ(s, t)≤ 1
is compact with respect to the norm ‖· , ·‖Vq′µ ,[0,T ]
. We first note that, for all (Y, Y ′) ∈ AT ,
‖Y ′‖q,[0,T ]≤ cµ(0, T )1q , ‖Y ′‖∞,[0,T ]≤M + cµ(0, T )
1q and ‖RY ‖r,[0,T ]≤ cµ(0, T )
1r ,
where the second bound follows from the fact that |Y ′t |≤ |Y ′0 |+|Y ′0,t|≤ M + ‖Y ′‖q,[0,T ]. The
p-variation of Y can also be controlled as follows. From Ys,t = Y ′sµs,t +RYs,t, we have
|Ys,t|p≤ 2p−1(‖Y ′‖p∞,[0,T ]|µs,t|p+|RYs,t|p), (s, t) ∈ ∆[0,T ],
and hence
‖Y ‖p,[0,T ]≤ 2p−1p (‖Y ′‖∞,[0,T ]‖µ‖p,[0,T ]+‖RY ‖p,[0,T ]) ≤ 2
p−1p (‖Y ′‖∞,[0,T ]‖µ‖p,[0,T ]+‖RY ‖r,[0,T ]),
since r < p (see e.g. [CG98, Remark 2.5]), and thus
‖Y ‖∞,[0,T ]≤M + ‖Y ‖p,[0,T ]≤M + 2p−1p
((M + cµ(0, T )
1q )‖µ‖p,[0,T ]+cµ(0, T )
1r
).
Therefore, by [FV10, Proposition 5.28], every sequence (Y n, Y n,′)n≥1 ⊂ AT has a convergentsubsequence, which we still denote by (Y n, Y n,′)n≥1, and limits Y ∈ Cp-var([0, T ];Rd) and Y ′ ∈
20 ALLAN, CUCHIERO, LIU, AND PROMEL
Cq-var([0, T ];Rd), such that |Y n0 −Y0|+‖Y n−Y ‖p′,[0,T ]→ 0 and |Y n,′
0 −Y ′0 |+‖Y n,′−Y ′‖q′,[0,T ]→ 0respectively as n→∞, for an arbitrary p′ > p. Since
|RY ns,t −RYn+1
s,t | ≤ |Y n,′s µs,t − Y n+1,′
s µs,t|+|Y ns,t − Y n+1
s,t |≤ |Y n,′
s − Y n+1,′s ||µs,t|+|Y n
t − Y n+1t |+|Y n
s − Y n+1s | −→ 0
as n→∞, uniformly in (s, t) ∈ ∆[0,T ], we have that
‖RY n −RY n+1‖r′,[0,T ] ≤ ‖RYn −RY n+1‖
rr′r,[0,T ] sup
(s,t)∈∆[0,T ]
|RY ns,t −RYn+1
s,t |r′−rr′
≤ 2rr′ cµ(0, T )
1r′ sup
(s,t)∈∆[0,T ]
|RY ns,t −RYn+1
s,t |r′−rr′ −→ 0
as n→∞. Thus, RYn
also converges to some RY in r′-variation.To see that the limit (Y, Y ′) ∈ AT , we simply note that
|Y ′s,t|q
cµ(s, t)+|RYu,v|r
cµ(u, v)= lim
n→∞
( |Y n,′s,t |q
cµ(s, t)+|RY nu,v |r
cµ(u, v)
)≤ 1,
and then take the supremum over (s, t) ∈ ∆[0,T ] and (u, v) ∈ ∆[0,T ] on the left-hand side.
Thus, AT is compact with respect to pµ,q′
T , and A is then compact in the topology generated
by the seminorms pµ,q′
T for T ∈ N.
Step 2: Now suppose that (πn, πn,′)n∈N, is a sequence of portfolios in AM,q(cµ). Corre-
spondingly, (πnµ , (πn
µ )′)n∈N, is then a sequence in A which, by the result in Step 1 above,
admits a convergent subsequence with respect to the seminorms ‖· , ·‖Vq′µ ,[0,T ]for T ∈ N.
Since ‖πµ , (πµ)′‖Vq′µ ,[0,T ]
= pµ,q′
T ((π, π′)), the convergence also applies to the corresponding sub-
sequence of (πn, πn,′)n∈N with respect to the seminorms pµ,q′
T T∈N. Let (φ, φ′) be the limit
of (the convergent subsequence of) (πnµ , (πn
µ )′)n∈N. It is then easy to see that φµ, the prod-
uct of controlled paths (φ, φ′) and (µ, I), is a cluster point of (πn, πn,′)n∈N in AM,q(cµ) with
respect to the seminorms pµ,q′
T T∈N.
In the next auxiliary result, we establish continuity of the relative wealth of admissibleportfolios with respect to the market portfolio. To this end, we recall the family of seminorms
pµ,q′
T T>0, defined in (4.2), and, for a given sequence β = βNN∈N with βN > 0 for all
N ∈ N and limN→∞ βN =∞, we introduce a metric dβ on AM,q(cµ) via
(4.6) dβ((π, π′), (φ, φ′)) := supN≥1
1
βNγNpµ,q
′
N ((π, π′)− (φ, φ′)),
where
γN := 1 +M + cµ(0, N)1q + cµ(0, N)
1r .
Since pµ,q′
N ((π, π′)) ≤ γN , we have that dβ((π, π′), (φ, φ′)) <∞ for all portfolios (π, π′), (φ, φ′) ∈AM,q(cµ). The metric dβ is thus well-defined on AM,q(cµ). Moreover, it is not hard to see thatthe topology induced by the metric dβ coincides with the topology generated by the family of
seminorms pµ,q′
T T∈N, so that (AM,q(cµ), dβ) is a compact metric space. For T > 0, we also
PORTFOLIO THEORY WITH ROUGH PATHS 21
denote
(4.7) ξT := ‖µ‖p,[0,T ]+‖Aµ‖ p2,[0,T ]+
d∑i=1
[µ]iiT .
Lemma 4.7. For any T > 0, we have that the estimate
(4.8) |log V πT − log V φ
T |≤ CβNγ2NξT dβ((π, π′), (φ, φ′))
holds for all (π, π′), (φ, φ′) ∈ AM,q(cµ), for some constant C which depends only on p, q′, r′
and the dimension d, where N = dT e, and V π denotes the relative wealth process as definedin (3.4). In particular, the map from AM,q(cµ)→ R given by (π, π′) 7→ V π
T is continuous withrespect to the metric dβ.
Proof. By Proposition 3.9 and the relation in (3.14), we have that, for any (π, π′) ∈ AM,q(cµ),
log V πT =
∫ T
0
πsµs
dµs −1
2
d∑i,j=1
∫ T
0
πisπjs
µisµjs
d[µ]ijs ,
which implies that, for (π, π′), (φ, φ′) ∈ AM,q(cµ),
|log V πT − log V φ
T |≤∣∣∣∣ ∫ T
0
πs − φsµs
dµs
∣∣∣∣+1
2
∣∣∣∣ d∑i,j=1
∫ T
0
(πis − φis)(πjs + φjs)
µisµjs
d[µ]ijs
∣∣∣∣.We aim to bound the two terms on the right-hand side. Let Aµ be the canonical rough
path lift of µ (as defined in Section A.3), namely Aµs,t =∫ ts µs,u ⊗ dµu. Writing N = dT e, by
the estimate for rough integrals in (2.3), we obtain∣∣∣∣ ∫ T
0
πs − φsµs
dµs
∣∣∣∣ . ‖Rπ−φµ ‖r′,[0,T ]‖µ‖p,[0,T ]+
∥∥∥(π − φµ
)′∥∥∥q′,[0,T ]
‖Aµ‖ p2,[0,T ]
+∣∣∣π0 − φ0
µ0
∣∣∣‖µ‖p,[0,T ]+∣∣∣(π − φ
µ
)′0
∣∣∣‖Aµ‖ p2,[0,T ]
. pµ,q′
N ((π, π′)− (φ, φ′))(‖µ‖p,[0,T ]+‖Aµ‖ p2,[0,T ])
≤ βNγNdβ((π, π′), (φ, φ′))(‖µ‖p,[0,T ]+‖Aµ‖ p2,[0,T ]).
For the second term, we note that
(4.9)
∣∣∣∣ ∫ T
0
(πis − φis)(πjs + φjs)
µisµjs
d[µ]ijs
∣∣∣∣ . ∥∥∥π − φµ ∥∥∥∞,[0,T ]
∥∥∥π + φ
µ
∥∥∥∞,[0,T ]
d∑i=1
[µ]iiT .
It follows from the relation πtµt
= π0µ0
+ (πµ)′0µ0,t +Rπµ
0,t, and the fact that µ takes values in the
bounded set ∆d+, that ∥∥∥π
µ
∥∥∥∞,[0,T ]
.M + cµ(0, T )1r ≤ γN .
It follows similarly from πt−φtµt
= π0−φ0µ0
+ (π−φµ )′0µ0,t +Rπ−φµ
0,t , that∥∥∥π − φµ
∥∥∥∞,[0,T ]
. pµ,q′
N ((π, π′)− (φ, φ′)) ≤ βNγNdβ((π, π′), (φ, φ′)).
22 ALLAN, CUCHIERO, LIU, AND PROMEL
Substituting back into (4.9), we obtain∣∣∣∣ ∫ T
0
(πis − φis)(πjs + φjs)
µisµjs
d[µ]ijs
∣∣∣∣ . βNγ2Ndβ((π, π′), (φ, φ′))
d∑i=1
[µ]iiT .
Combining the inequalities above, we deduce the desired estimate.
In the following, we will sometimes write simply AM,q := AM,q(cµ) for brevity.
For (π, π′) ∈ AM,q, we have by definition that π is a µ-controlled path. We also havethat the relative wealth V π is also a µ-controlled rough path—as can be seen for instancefrom Proposition 3.9—and hence the product πV π is also a controlled path. Let ν be a fixedprobability measure on (AM,q, dβ). Observe that for every T > 0 the space Vqµ([0, T ];Rd) ofcontrolled paths is a Banach space, and that, as we will see during the proof of Lemma 4.8below, V π is the unique solution to the rough differential equation (4.11), which implies thatthe mapping π 7→ V π|[0,T ]∈ V
qµ([0, T ];Rd) is continuous by the continuity of the Ito–Lyons
map (see e.g. [Lej12, Theorem 1]). Hence, for every T > 0 we can define the Bochner integral∫AM,q(πV
π)|[0,T ] dν(π), which is thus itself another controlled path defined on [0, T ]. Theµ-controlled path
(4.10) πνt :=
∫AM,q πtV
πt dν(π)∫
AM,q Vπt dν(π)
, t ∈ [0,∞),
is then well-defined, and defines indeed a portfolio in Vqµ, called the universal portfolio asso-ciated to the set AM,q of admissible portfolios.
Lemma 4.8. Let πν be the universal portfolio as defined in (4.10). Then, for all T > 0,
V πν
T =
∫AM,q
V πT dν(π).
Proof. By Proposition 3.9 and the relation in (3.14), we have, for any portfolio π,
V πt = exp
(∫ t
0
πsµs
dµs −1
2
d∑i,j=1
∫ t
0
πisπjs
µisµjs
d[µ]ijs
).
Setting Z :=∫ ·
0πsµs
dµs, by Lemma B.1, we can rewrite the relation above as V π = exp(Z −12 [Z]). Thus, by Lemma A.5, Lemma A.4 and Proposition A.2, we deduce that V π is theunique solution Y to the linear rough differential equation
(4.11) Yt = 1 +
∫ t
0Ysπsµs
dµs, t ≥ 0.
It is therefore sufficient to show that the path t 7→∫AM,q V
πt dν(π) also satisfies the RDE
(4.11) with π replaced by πν . By the definition of the universal portfolio in (4.10), we have
(4.12)
∫AM,q
V πs dν(π)
πνsµs
=
∫AM,q
πsµsV πs dν(π).
Recalling that V π satisfies (4.11), we know that
V πt = 1 +
∫ t
0
πsµsV πs dµs.
PORTFOLIO THEORY WITH ROUGH PATHS 23
By the Fubini theorem for rough integration—namely Theorem A.6—we then have that∫AM,q
V πt dν(π) = 1 +
∫ t
0
∫AM,q
πsµsV πs dν(π) dµs
= 1 +
∫ t
0
∫AM,q
V πs dν(π)
πνsµs
dµs,
where we used (4.12) to obtain the last line. Hence, both V πν and∫AM,q V
π dν(π) are theunique solution of the same RDE, and thus coincide.
With these preparations in place, we now aim to compare the growth rates of the universalportfolio (4.10) and the best retrospectively chosen portfolio. For this purpose, we fix anM > 0, and assume that there exists a compact metric space (K, dK) together with a mappingι : (K, dK)→ (AM,q, dβ) such that ι is continuous and injective (and thus a homeomorphismonto its image), and that for every T > 0 and x, y ∈ K, we have that
(4.13) |log Vι(x)T − log V
ι(y)T |≤ Cλ(T )dK(x, y),
where λ is a positive function of T , and C is a universal constant independent of T . Here welist some examples of (K, dK), ι and λ:
(1) K = Cp+α,K(∆d+;Rd) = G ∈ Cp+α(∆
d+;Rd) : ‖G‖Cp+α≤ K, G ≥ 1
K , dK(G, G) =
‖G − G‖C2 , ι(G) = πG, where α > 0 and πG is a classical functionally generatedportfolio of the form (3.11). In this case we can take λ(T ) = 1 + maxi=1,...,d[µ]iiT ; seethe proof of [CSW19, Lemma 4.4].
(2) K = C2+α,K(∆d+;Rd) = F ∈ C2+α(∆
d+;Rd) : ‖F‖C2+α≤ K, dK(F, F ) = ‖F − F‖C2 ,
ι(F ) = πF , where α ∈ (0, 1] and πF is a functionally controlled portfolio defined asin (4.4). In this case one may take λ(T ) = (1 + ‖µ‖2p,[0,T ])ξT , where ξT is defined in
(4.7); see Lemma 4.11 below.(3) K = AM,q, dK = dβ, ι = IdAM,q . In view of (4.8) we have λ(T ) = βdT eγ
2dT eξT .
Given such a compact space (K, dK) equipped with an embedding ι as above, we define
V ∗,K,ιT = supx∈K
Vι(x)T = sup
π∈ι(K)V πT .
By the compactness of K and the continuity provided by the estimate in (4.13), we have that,for each T > 0, there exists a portfolio π∗,T ∈ ι(K), which can be expressed as π∗,T = ι(x∗)for some x∗ ∈ K, known as the best retrospectively chosen portfolio associated with K and ι,such that
(4.14) V ∗,K,ιT = V π∗,T .
The following theorem provides an analogue of [CSW19, Theorem 4.11] in our rough pathsetting.
Theorem 4.9. Let (K, dK) be a compact metric space equipped with a continuous embeddingι : (K, dK)→ (AM,q, dβ) which satisfies the bound in (4.13) for some positive function λ. Letm be a probability measure on K with full support, and let ν = ι∗(m) denote the pushforwardmeasure on AM,q. If limT→∞ λ(T ) =∞, then
limT→∞
1
λ(T )
(log V ∗,K,ιT − log V πν
T
)= 0.
24 ALLAN, CUCHIERO, LIU, AND PROMEL
In particular, if K = Cp+α,K(∆d+;Rd) = G ∈ Cp+α(∆
d+;Rd) : ‖G‖Cp+α≤ K, G ≥ 1
K ,dK(G, G) = ‖G − G‖C2 , ι(G) = πG, where πG is a classical functionally generated portfolioof the form (3.11), and λ(T ) = 1 + maxi=1,...,d[µ]iiT , then one also infers the version of Cover’stheorem obtained in [CSW19, Theorem 4.11].
Proof of Theorem 4.9. As the inequality “≥” is trivial, we need only show the reverse inequal-ity. As K is compact and m has full support, we have that, for any η ∈ (0, 1), there exists aδ > 0 such that every η-ball around a point x ∈ K with respect to dK has m-measure biggerthan δ.
Let T > 0 be such that λ(T ) ≥ 1, and let π∗,T = ι(x∗) be the best retrospectively chosenportfolio, as in (4.14). For any portfolio π = ι(x) ∈ ι(K) ⊆ AM,q(cµ) such that dK(x, x∗) ≤ η,the estimate in (4.13) implies that
1
λ(T )
(log V π
T − log V π∗,TT
)≥ −CdK(x, x∗) ≥ −Cη,
for some constant C. For any ε > 0, we can therefore choose η small enough such that
(4.15)1
λ(T )
(log V π
T − log V π∗,TT
)≥ −ε.
Let Bη(x∗) denote the η-ball in K around the point x∗ with respect to the metric dK, which
has m-measure |Bη(x∗)|≥ δ. By Lemma 4.8 and Jensen’s inequality, we have that
(V πν
T )1
λ(T ) ≥(∫
Bη(x∗)Vι(x)T dm(x)
) 1λ(T )
≥ |Bη(x∗)|1
λ(T )−1∫Bη(x∗)
(Vι(x)T )
1λ(T ) dm(x).
Then, using (4.15), we have(V πν
T
V π∗,TT
) 1λ(T )
≥ |Bη(x∗)|1
λ(T )−1∫Bη(x∗)
(Vι(x)T
Vι(x∗)T
) 1λ(T )
dm(x) ≥ |Bη(x∗)|1
λ(T ) e−ε ≥ δ1
λ(T ) e−ε.
Taking ε > 0 arbitrarily small (which determines η and hence also δ) and then T > 0sufficiently large, we deduce the desired inequality.
4.3. Universal portfolios based on functionally controlled portfolios. The most fre-quently considered classes of portfolios are those which are generated by functions acting onthe underlying price trajectories, such as the functionally generated portfolios in Lemma 4.3.In this section we shall investigate the growth rate of universal portfolios based on the moregeneral class of functionally controlled portfolios, as introduced in Example 4.4. More pre-cisely, we fix constants α ∈ (0, 1] and K > 0, and consider the sets
C2+α,K(∆d+;Rd) := F ∈ C2+α(∆
d+;Rd) : ‖F‖C2+α≤ K
and
F2+α,K := (πF , πF,′) : F ∈ C2+α,K(∆d+;Rd),
where the portfolio πF is of the form in (4.4). Here we recall that C2+α denotes the space oftwice continuously differentiable functions whose second derivative is α-Holder continuous.
Lemma 4.10. For any T > 0 and any F,G ∈ C2+α,K(∆d+;Rd), we have that
(4.16) pµ,pT ((πF , πF,′)− (πG, πG,′)) ≤ C‖F −G‖C2(1 + ‖µ‖2p,[0,T ]),
PORTFOLIO THEORY WITH ROUGH PATHS 25
where the constant C depends only on p, d and K. Considering the map Φ:C2+α,K(∆d+;Rd)→
F2+α,K given by2
F 7→ Φ(F ) := (πF , πF,′),
where πF is of the form in (4.4), we thus have that Φ is continuous with respect to the C2-
distance on C2+α,K(∆d+;Rd) and each of the seminorms pµ,pT T>0 on F2+α,K ⊂ AM,p(cµ).
As the notation suggests, here pµ,pT is defined as in (4.2) with q′ replaced by p.
Proof. In the following, for notational simplicity we will omit the Gubinelli derivative in thenorms ‖· , ·‖Vpµ,[0,T ] and seminorms pµ,pT (( · , ·)); that is, we will write e.g. ‖π‖Vpµ,[0,T ] instead of
‖π, π′‖Vpµ,[0,T ]. Let F,G ∈ C2+α,K and s ≤ t. We have∣∣∣(DF −DG)(µt)− (DF −DG)(µs)∣∣∣ =
∣∣∣∣ ∫ 1
0(D2F −D2G)(µs + λµs,t)µs,t dλ
∣∣∣∣≤ ‖D2F −D2G‖∞|µs,t|,
so that
‖DF (µ)−DG(µ)‖p,[0,T ]≤ ‖F −G‖C2‖µ‖p,[0,T ].
Similarly, since
RF (µ)s,t = F (µt)− F (µs)−DF (µs)µs,t =
∫ 1
0
∫ 1
0D2F (µs + λ1λ2µs,t)µ
⊗2s,t λ1 dλ2 dλ1,
we have
‖RF (µ) −RG(µ)‖ p2,[0,T ]≤ ‖F −G‖C2‖µ‖2p,[0,T ].
Thus, for µ-controlled paths (F (µ),DF (µ)) and (G(µ),DG(µ)), we have that
(4.17) ‖F (µ)−G(µ)‖Vpµ,[0,T ]. ‖F −G‖C2(1 + ‖µ‖2p,[0,T ]).
Writing πFt /µt = F (µt) + (1 − µt · F (µt))1 and πGt /µt = G(µt) + (1 − µt · G(µt))1, we havethat
πFt − πGtµt
= F (µt)−G(µt)− (µt · (F (µt)−G(µt)))1,
so that
(4.18) pµ,pT (πF − πG) . ‖F (µ)−G(µ)‖Vpµ,[0,T ]+‖µ · (F (µ)−G(µ))‖Vpµ,[0,T ].
Similarly to the proof of Lemma 4.3, noting that Rµ·(F (µ)−G(µ))s,t = µs · RF (µ)−G(µ)
s,t + µs,t ·(F (µ)−G(µ))s,t, we have that
|Rµ·(F (µ)−G(µ))s,t |≤ ‖µ‖∞,[0,T ]|R
F (µ)−G(µ)s,t |+|µs,t||(F (µ)−G(µ))s,t|. ‖F −G‖C2 |µs,t|2,
where we used the fact that µ is bounded, and we deduce that
‖µ · (F (µ)−G(µ))‖Vpµ,[0,T ]. ‖F −G‖C2(1 + ‖µ‖2p,[0,T ]).
Combining this with (4.17) and (4.18), we obtain the estimate in (4.16), which then impliesthe desired continuity of Φ.
2Note that Φ plays the role of the embedding ι in the previous section.
26 ALLAN, CUCHIERO, LIU, AND PROMEL
Lemma 4.11. For any T > 0 and any F,G ∈ C2+α,K(∆d+;Rd), we have that
(4.19) |log V πF
T − log V πG
T |≤ C‖F −G‖C2(1 + ‖µ‖2p,[0,T ])ξT ,
where ξT is defined as in (4.7), and the constant C depends only on p, d and K.
Proof. We recall that during the proof of Lemma 4.7 we showed that
|log V πF
T − log V πG
T |≤∣∣∣∣ ∫ T
0
πFs − πGsµs
dµs
∣∣∣∣+1
2
∣∣∣∣ d∑i,j=1
∫ T
0
(πF,is − πG,is )(πF,js + πG,js )
µisµjs
d[µ]ijs
∣∣∣∣,and (in the current setting replacing q′ by p)∣∣∣∣ ∫ T
0
πFs − πGsµs
dµs
∣∣∣∣ . pµ,pT ((πF , πF,′)− (πG, πG,′))(‖µ‖p,[0,T ]+‖Aµ‖ p2,[0,T ]).
By the estimate in (4.16), we obtain∣∣∣∣ ∫ T
0
πFs − πGsµs
dµs
∣∣∣∣ . (1 + ‖µ‖2p,[0,T ])(‖µ‖p,[0,T ]+‖Aµ‖ p2,[0,T ])‖F −G‖C2 .
Since ‖F‖C2+α≤ K and ‖G‖C2+α≤ K, recalling (4.4), we can verify that∣∣∣∣(πF,is − πG,is )(πF,js + πG,js )
µisµjs
∣∣∣∣ . ‖F −G‖C2 .
Hence, we have that∣∣∣∣ d∑i,j=1
∫ T
0
(πF,is − πG,is )(πF,js + πG,js )
µisµjs
d[µ]ijs
∣∣∣∣ . ‖F −G‖C2
d∑i=1
[µ]iiT .
Combining the estimates above, we obtain (4.19).
As a special case of Theorem 4.9, we can deduce an asymptotic growth rate for the universalportfolio in the case that our portfolios are restricted to the class F2+α,K of functionallycontrolled portfolios.
Let m be a fixed probability measure on C2+α,K = C2+α,K(∆d+;Rd), and define ν := Φ∗m
as the pushforward measure on F2+α,K of m under the map Φ given in Lemma 4.10. Theuniversal portfolio based on functionally controlled portfolios is then defined by
(4.20) πνt :=
∫F2+α,K πtV
πt dν(π)∫
F2+α,K V πt dν(π)
, t ∈ [0,∞),
and the wealth process of the best retrospectively chosen portfolio is defined as
(4.21) V ∗,K,αT := supπ∈F2+α,K
V πT = sup
F∈C2+α,K
V πF
T .
By Lemma 4.11, the mapping F 7→ V πF
T is a continuous map on C2+α,K with respect to theC2-norm. We also have that C2+α,K is compact with respect to the C2-norm (see [CSW19,Lemma 4.1]). Combining these two facts, we see that, for each T > 0, there exists a functionF ∗T ∈ C2+α,K such that
V ∗,K,αT = V πF∗T
T .
PORTFOLIO THEORY WITH ROUGH PATHS 27
The following result follows from Theorem 4.9 applied with K = C2+α,K , dK(F,G) =‖F −G‖C2 , ι = Φ and λ(T ) = (1 + ‖µ‖2p,[0,T ])ξT .
Theorem 4.12. Suppose that limT→∞(1 +‖µ‖2p,[0,T ])ξT =∞, where as usual ξT is defined as
in (4.7). Let m be a probability measure on C2+α,K with full support. Let πν be the universal
portfolio as defined in (4.20), and define V ∗,K,αT as in (4.21). Then
limT→∞
1
(1 + ‖µ‖2p,[0,T ])ξT
(log V ∗,K,αT − log V πν
T
)= 0.
Remark 4.13. Let us briefly comment on the “clock” (1+‖µ‖2p,[0,T ])ξT in Theorem 4.12. For
intuition, let us suppose that µ is a d-dimensional Brownian motion, and that Aµ is the Ito
rough path lift of µ. In this case we have that∑d
i=1[µ]iiT = Td. Moreover, using the enhancedBDG inequalities ([FV10, Theorem 14.12]), we have that
E[‖µ‖2p,[0,T ]+‖A
µ‖ p2,[0,T ]
]∼ E
[ d∑i=1
[µ]iiT
]= Td.
We infer that, “on average”, ξT ≤ 1 + ‖µ‖2p,[0,T ]+‖Aµ‖ p
2,[0,T ]+
∑di=1[µ]iiT ∼ 1 + Td. The
function ξT thus satisfies, in some sense, lim supT→∞ξTT < ∞, which has the same form as
the condition in [CSW19, (58)]. The additional term 1 + ‖µ‖2p,[0,T ] comes from the structure
of controlled paths, and from the enhanced BDG inequalities, we also have that
E[(1 + ‖µ‖2p,[0,T ])ξT ] . E[(1 + ‖µ‖2p,[0,T ]+‖A
µ‖ p2,[0,T ])
2]
+ E[(1 + ‖µ‖2p,[0,T ])
d∑i=1
[µ]iiT
]
. 1 + E[( d∑
i=1
[µ]iiT
)2]= 1 + T 2d2.
Theorem 4.12 therefore suggests informally that by using functionally controlled portfolios,the asymptotic growth rate is at most O(T 2), i.e.
limT→∞
1
T 2
(log V ∗,K,αT − log V πν
T
)= 0,
while the corresponding rate for functionally generated portfolios is O(T ) ∼ O(∑d
i=1[µ]iiT ).
Remark 4.14. Strictly speaking, Theorems 4.9 (which also recovers the version of Cover’stheorem established in [CSW19]) and 4.12 do not say that the universal portfolio πν performsasymptotically as well as the best retrospectively chosen one; rather, they provide boundson how large the gap can become as time increases. For instance, for classical functionallygenerated portfolios of form (3.11) the gap is o(maxi=1,...,d[µ]iiT ), and for functionally controlledportfolios of form (4.4) the gap is o((1 + ‖µ‖2p,[0,T ])ξT ).
Remark 4.15. We infer that the asymptotic discrepancy between the wealth generated bythe best retrospectively chosen portfolio and the universal portfolio is typically larger than ina probabilistic setup. This is not surprising, and is fundamentally because we are perform-ing a worst-case analysis, whereby we search through a space of market portfolio trajectoriesconsisting of essentially every possible path of finite p-variation (or at least those which sat-isfy Property (RIE)). This is a typical situation in model-free finance and rough path theory,
28 ALLAN, CUCHIERO, LIU, AND PROMEL
and is an inevitable consequence of adopting such a robust approach. On the other hand, aswe will see in Section 5, if the market portfolio is realized by a stochastic process, then onecan improve the pathwise growth rate (1 + ‖µ‖2p,[0,T ])ξT for functionally controlled portfolios
to the growth rate T (which was established in [CSW19] for classical functionally generatedportfolios). Note however that this only applies to market portfolios µ outside of a null set.
4.4. The non-triviality of the asymptotic growth rate. In this section we will show thatthe asymptotic growth rate λ(T ) = (1 + ‖µ‖2p,[0,T ])ξT for functionally controlled portfolios, as
established in Theorem 4.12, is non-trivial, in the sense that there exists an instance of themarket portfolio µ = (µt)t∈[0,∞) such that
lim supT→∞
log V ∗,K,αT
(1 + ‖µ‖2p,[0,T ])ξT> 0 and lim
T→∞
1
(1 + ‖µ‖2p,[0,T ])ξT
(log V ∗,K,αT − log V πν
T
)= 0,
where ν = Φ∗m for an arbitrary probability measure m on C2+α,K with full support.
Lemma 4.16. Let p ∈ (2, 3) as usual, and then fix λ > 0 such that 1p < λ < 1
2 . Let d = 3
and let µ = (µt)t∈[0,∞) be the continuous ∆3+-valued path given by
µt =
µ1t
µ2t
µ3t
=
13(1 + k−λ
3 (1− cos t))13(1 + k−λ
3 sin t)13(1 + k−λ
3 (cos t− 1− sin t))
, t ∈ [2π(k − 1), 2πk),
for each k ∈ N. For α ∈ (0, 1] and K > 0, let V ∗,K,αT be the wealth induced by the best
retrospectively chosen portfolio over F2+α,K at time T . Then
lim supT→∞
log V ∗,K,αT
(1 + ‖µ‖2p,[0,T ])ξT> 0.
Proof. Recall that for any portfolio π, it follows from Proposition 3.9 that log V πT =
∫ T0
πsµs
dµs−12
∑di,j=1
∫ T0
πisπjs
µisµjs
d[µ]ijs . Clearly, since µ is continuous with bounded variation on every com-
pact interval, we have that [µ] = 0, so that the second term vanishes. For any functionallycontrolled portfolio πF ∈ F2+α,K , using the relation
πF,it
µit= F i(µt) + 1−
d∑j=1
µjtFj(µt), i = 1, . . . , d,
together with the fact that∑d
i=1 dµit = 0 (since∑d
i=1 µit = 1), we deduce that
(4.22) log V πF
T =
∫ T
0
πFtµt
dµt =
d∑i=1
∫ T
0
πF,it
µitdµit =
d∑i=1
∫ T
0F i(µt) dµit.
We now choose the function F ∈ C2+α,K given by
F (x) =
x2
00
PORTFOLIO THEORY WITH ROUGH PATHS 29
for x = (x1, x2, x3)> ∈ ∆3+. Substituting this function into (4.22), we have
log V πF
T =
∫ T
0
πFtµt
dµt =3∑i=1
∫ T
0F i(µt) dµit =
∫ T
0µ2t dµ1
t .
For n ∈ N, we compute∫ 2πn
0µ2t dµ1
t =
n∑k=1
∫ 2πk
2π(k−1)µ2t dµ1
t =n∑k=1
∫ 2π
0
1
3
(1 +
k−λ
3sin t
)· k−λ
9sin tdt
=
n∑k=1
k−2λ
81
∫ 2π
0sin2 tdt =
π
81
n∑k=1
k−2λ,
and note that
‖µ‖p,[0,2πn].
( n∑k=1
k−λp)1p
<
( ∞∑k=1
k−λp)1p
<∞
for every n ∈ N.
Writing Aµs,t = [Aµ,i,js,t ]i,j=1,2,3 =∫ ts (µu − µs) ⊗ dµu for the canonical rough path lift of µ,
and using the monotonicity of trigonometric functions on the intervals [0, π2 ], [π2 , π], [π, 3π2 ]
and [3π2 , 2π], one can readily check that
‖Aµ‖ p2,[0,2πn]∼ A
µ,2,10,2πn =
∫ 2πn
0µ2t dµ1
t ∼n∑k=1
k−2λ.
Recalling that ξT = ‖µ‖p,[0,T ]+‖Aµ‖ p2,[0,T ] (since [µ] = 0), and combining the calculations
above, we deduce that, for T = 2πn,
log V ∗,K,αT
(1 + ‖µ‖2p,[0,T ])ξT≥ log V πF
2πn
(1 + ‖µ‖2p,[0,2πn])ξ2πn&
∑nk=1 k
−2λ
1 +∑n
k=1 k−2λ
−→ 1 as n→∞,
where we used the fact that 2λ < 1.
The example in Lemma 4.16 thus shows that for functionally controlled portfolios πF
generated by a function F ∈ C2+α,K which is not necessarily of gradient-type, the asymptoticgrowth rate (1 + ‖µ‖2p,[0,T ])ξT appearing in Theorem 4.12 is actually sharp, in the sense that
the log-relative wealth log V πF
T and the rate (1 + ‖µ‖2p,[0,T ])ξT grow at the same rate (up to a
multiplicative constant) as T →∞.
4.5. Functionally controlled portfolios have better performance. Let us conclude thissection by showing that classical functionally generated portfolios of form in (3.11), whichare induced by functions of gradient type, are in general not optimal among the class offunctionally controlled portfolios of the form in (4.4).
Let µ be a continuous ∆d+-valued path which, for simplicity, we assume to have finite
variation on every bounded interval (and which therefore trivially satisfies Property (RIE)).
For any F ∈ C2+α,K(∆d+;Rd), we know, as we saw in (4.22) above, that for every T > 0,
log V πF
T =
∫ T
0
πFsµs
dµs −1
2
d∑i,j=1
∫ T
0
πF,is πF,js
µisµjs
d[µ]ijs =
∫ T
0F (µs) dµs,
30 ALLAN, CUCHIERO, LIU, AND PROMEL
since the quadratic variation [µ] vanishes. Suppose now that the generating function F wereof gradient-type, so that F = ∇f for some suitably smooth real-valued function f . We thenhave that
log V πF
T =
∫ T
0∇f(µs) dµs = f(µT )− f(µ0),
which implies together with the mean value theorem that
|log V πF
T |≤ ‖∇f‖∞|µT − µ0|= ‖F‖∞|µT − µ0|≤ 2K,
as ‖F‖∞≤ K and µT , µ0 ∈ ∆d+. In particular, we have that
(4.23) supT≥0
log V πF
T ≤ 2K <∞
for every generating function F of gradient type.
Now let µ be the market portfolio given in Lemma 4.16, and let F (x1, x2, x3) = (x2, 0, 0)>,which we note is not of gradient type. In the proof of Lemma 4.16 we saw, for T = 2πn
with any n ∈ N, that log V πF
T =∫ T
0 µ2t dµ1
t = π81
∑nk=1 k
−2λ for some positive λ < 12 . We thus
immediately have that
(4.24) lim supT→∞
log V πF
T =∞.
Comparing (4.24) with (4.23), it is clear that the best retrospectively chosen portfolio overthe set of functionally controlled portfolios cannot be of gradient type. Indeed, we infer thatamong the class of all functionally controlled portfolios, those corresponding to gradient-typegenerating functions are in general far from being optimal, demonstrating the need to gobeyond gradient-type generating functions.
5. Functionally controlled portfolios in probabilistic models
In this section we shall demonstrate some further links between our purely pathwise theoryand classical stochastic portfolio theory in a probabilistic setting. In particular, this willallow us to illustrate again the advantages of functionally controlled portfolios, as introducedin Example 4.4, compared to (pathwise) functionally generated portfolios (see Lemma 4.3),as were previously treated in [SSV18, CSW19] based on Follmer integration.
5.1. Probabilistic model for the market portfolio. Whereas in the previous sections weworked in a purely pathwise setting, we now assume that the market portfolio (also knownas the market weights process) µ = (µ1
t , . . . , µdt )t∈[0,∞) is described by a time-homogeneous
Markovian Ito-diffusion with values in ∆d+, of the form
µt = µ0 +
∫ t
0c(µs)λ(µs) ds+
∫ t
0
√c(µs) dWs, t ∈ [0,∞),(5.1)
where µ0 is distributed according to some measure ρ on ∆d+, W is a d-dimensional Brownian
motion and√· denotes the matrix square root. We assume that µ is the canonical process
defined on path space (Ω,F ,P), i.e. Ω = C([0,∞); ∆d+), F = σ(µt : t ∈ [0,∞)), and P denotes
the law of µ. For the moment λ is just assumed to be a Borel measurable function from ∆d+
to Rd. Writing Sd+ for the set of positive semi-definite symmetric matrices, c ∈ C(∆d+;Sd+) is
such that
c(x)1 = 0 for all x ∈ ∆d+.
PORTFOLIO THEORY WITH ROUGH PATHS 31
The latter requirement is necessary to guarantee that the process µ lies in ∆d+. For a complete
characterization of stochastic invariance of the closed simplex (under additional regularity con-ditions on the coefficients λ and c) we refer to [ABI19, Theorem 2.3] and the references therein.To ensure that the process stays in the open simplex ∆d
+, conditions for non-attainment ofthe boundary are established for instance in [FL16, Theorem 5.7]. These conditions build onversions of what is sometimes called “McKean’s argument” (see [MPS11] for an overview andfurther references).
We further suppose that the so-called structure condition is satisfied, that is
(5.2)
∫ T
0λ>(µs)c(µs)λ(µs) ds <∞ P-a.s., for all T ∈ [0,∞),
which is equivalent to “no unbounded profit with bounded risk” (NUPBR); see e.g. Theo-rem 3.4 in [HS10].
Remark 5.1. As (NUPBR) is satisfied due to (5.2), the sample paths of µ almost surelysatisfy Property (RIE) with respect to every p ∈ (2, 3) and a suitable sequence of partitions,cf. Remark 2.8.
We further impose the following ergodicity assumption in spirit of [Ebe16, Section 2.2,Theorem 2.6 and Section 2.2.3, Theorem 2.8] along with an integrability condition on λ.
Assumption 5.2. We assume that the market portfolio µ, given by the dynamics in (5.1),is an ergodic process with stationary measure ρ on ∆d
+. That is, we suppose that ρpt = ρ forevery t ∈ [0,∞), where here (pt)t∈[0,∞) denotes the transition probability of µ. Furthermore,
we suppose that λ ∈ L2(∆d+, ρ;Rd).
Note that the assumption that ρ is a stationary measure implies that the shift semigroupΘt(ω) = ω(t+ ·), t ∈ [0,∞), ω ∈ Ω, preserves the measure P, in the sense that P Θ−1
t = P.Hence, the “ergodic theorem in continuous time” (see [Ebe16, Section 2.2, Theorem 2.6,Theorem 2.8]) can be applied.
While on the pathwise market Ωp the portfolios were given by µ-controlled paths (π, π′) ∈Vqµ (recall Definition 2.3), in the current semimartingale setting we consider a portfolio π tobe an element of the set Π of all predictable processes π taking values in ∆d, such that theIto integral ∫ T
0
πsµs
dµs =
∫ T
0
d∑i=1
πisµis
dµis
is well-defined for every T ∈ [0,∞). As established in [CSW19, Section 4.2.3], for π ∈ Π, therelative wealth process (recall (3.4)) can be written in the usual form, that is
V πT = exp
(∫ T
0
πsµs
dµs −1
2
∫ T
0
d∑i,j=1
πisπjs
µisµjs
cij(µs) ds
), T ∈ [0,∞).(5.3)
Remark 5.3. Note that if (π, π′) is an adapted process with sample paths which are almostsurely µ-controlled paths, then it is predictable, and under Property (RIE) the rough integral
interpretation of∫ T
0πsµs
dµs coincides almost surely with the Ito integral interpretation. Indeed,
the rough integral can be approximated by left-point Riemann sums (see Theorem 2.12), whilethe Ito integral can be approximated by the same Riemann sums in probability (see e.g. [Pro04,Theorem II.21]). Moreover, as established in Proposition 3.9, the identity in (5.3) holds evenin a pathwise setting.
32 ALLAN, CUCHIERO, LIU, AND PROMEL
5.2. Log-optimal portfolios and their asymptotic growth rates. For a given T > 0,the log-optimal portfolio π is the maximizer of the optimization problem
(5.4) supπ∈Π
E[log V πT ].
We write
VT := V πT
for the corresponding wealth process. As shown in [CSW19, Section 4.2.3], if µ satisfies thedynamics in (5.1), then π = (π1, . . . , πd) can be expressed as
πit = µit
(λi(µt) + 1−
d∑j=1
µjtλj(µt)
), t ∈ [0,∞),(5.5)
and, due to (5.3), the expected value of the log-optimal portfolio satisfies
E[log VT ] = supπ∈Π
E[log V πT ] =
1
2E[ ∫ T
0λ>(µs)c(µs)λ(µs) ds
].(5.6)
We suppose that the log-optimal portfolio has finite maximal expected utility and requirethus additionally to (5.2) that
E[ ∫ T
0λ>(µs)c(µs)λ(µs) ds
]<∞.
From the expression in (5.5), we see immediately that the log-optimal portfolio π belongsto the class of functionally controlled portfolios, as defined in Example 4.4, whenever λ issufficiently smooth. In general, however, it does not belong to the smaller class of functionallygenerated portfolios, as we will see in Section 5.3.
In (5.4), the supremum is taken over all predictable strategies in Π. However, since theoptimizer is actually of the form in (5.5), we can also take the supremum in (5.4) over asmaller set. Indeed, it is sufficient to consider (functionally controlled) portfolios of the form
(πFt )i = µit
(F i(µt) + 1−
d∑j=1
µjtFj(µt)
),(5.7)
for functions F in the space L2(∆d+, ρ;Rd) due to Assumption 5.2.
Clearly, any portfolio πF of the form in (5.7) can itself be considered as a function πF ∈L2(∆d
+, ρ;Rd) which maps x 7→ πF (x), where
(5.8) [πF (x)]i = xi(F i(x) + 1−
d∑j=1
xjF j(x)
),
with the corresponding portfolio then being given by t 7→ πF (µt).In the current probabilistic setting we establish the following equivalence of the asymptotic
growth rates of the log-optimal and the universal portfolio based on functionally controlledportfolios of the form in (5.7).
Theorem 5.4. Let µ be a market weights process with the dynamics in (5.1) satisfying As-
sumption 5.2. Suppose that λ ∈ L2(∆d+, ρ;Rd) and c ∈ C(∆
d+;Sd+). Let m be a probability
PORTFOLIO THEORY WITH ROUGH PATHS 33
measure on L2(∆d+, ρ;Rd) such that λ ∈ supp(m). Define the universal portfolio πν analo-
gously to (4.20) but with ν now being the pushforward measure of m under the mapping
F 7→ πF
with πF as in (5.8), cf. [CSW19, Section 4.2.2]). Suppose that there exists an integrablerandom variable w such that, for each T > 0, the growth rate of the universal portfoliosatisfies
(5.9)1
Tlog V πν
T ≥ −w.
We then have that
lim infT→∞
1
Tlog V πν
T = limT→∞
1
Tlog VT = L P-a.s.
where L is given by
L :=1
2
∫∆d
+
λ>(x)c(x)λ(x) ρ(dx).
As preparation for the proof of Theorem 5.4, we need the following technical lemma, whichis an adaptation of [HKT01, Lemma 3.1].
Lemma 5.5. Let (fn)n∈N be a sequence of non-negative measurable functions on some topo-logical space A, such that the map a 7→ lim infn→∞ fn(a) is continuous at some point a ∈ A.Let ν be a probability measure on A with a ∈ supp(ν). Then
lim infn→∞
fn(a) ≤ lim infn→∞
(∫Afnn (a) ν(da)
) 1n
.
Proof. Let g ≥ 0 be bounded measurable function such that∫A g(a) ν(da) = 1. By Fatou’s
Lemma and Holder’s inequality,∫A
lim infn→∞
fn(a)g(a) ν(da) ≤ lim infn→∞
∫Afn(a)g(a) ν(da)
≤ lim infn→∞
(∫Afnn (a) ν(da)
) 1n(∫Ag
nn−1 (a) ν(da)
)n−1n
= lim infn→∞
(∫Afnn (a) ν(da)
) 1n
where the last line follows from the fact that limn→∞∫A g
nn−1 ν(da) =
∫A g(a) ν(da) by the
dominated convergence theorem. Since g was arbitrary, a lies in the support of ν, andlim infn→∞ fn is continuous at a, we deduce the result.
Proof of Theorem 5.4. By the conditions on λ and c, and the fact that we consider portfo-lios of the form in (5.7) with F ∈ L2(∆d
+, ρ;Rd), we see that the assumptions of [CSW19,
Theorem 4.9] are satisfied. Thus, for each F ∈ L2(∆d+, ρ;Rd), we have that
(5.10) limT→∞
1
Tlog V πF
T = LπF
P-a.s.,
where
LπF
:=
∫∆d
+
(πF (x)
x
)>c(x)λ(x) ρ(dx)− 1
2
∫∆d
+
(πF (x)
x
)>c(x)
(πF (x)
x
)ρ(dx).
34 ALLAN, CUCHIERO, LIU, AND PROMEL
Taking the supremum over F ∈ L2(∆d+, ρ;Rd), we find that
supF∈L2(∆d
+,ρ;Rd)
LπF
= Lπλ
= L.
Recalling (5.5) and (5.10), it follows that, P-a.s.,
limT→∞
1
Tlog VT = Lπ
λ= L.
Note that the map
F 7→ exp(LπF
) = limT→∞
(V πF
T
) 1T
is continuous with respect to the L2(∆d+, ρ;Rd)-norm. Thus, recalling Lemma 4.8, and ap-
plying Lemma 5.5 with fT (F ) = (V πF
T )1T , we deduce that
(5.11) L = limT→∞
1
Tlog VT ≤ lim inf
T→∞
1
Tlog V πν
T P-a.s.
On the other hand, by the definition of the log-optimal portfolio,
(5.12) E[log V πν
T ] ≤ E[log VT ].
By (5.6) and the ergodicity of the process µ, we have that
(5.13) limT→∞
1
TE[log VT ] = L.
By Fatou’s lemma (which we may apply by the condition in (5.9)), (5.12), (5.13) and (5.11),we then have that, P-a.s.,
E[
lim infT→∞
1
Tlog V πν
T
]≤ lim inf
T→∞
1
TE[log V πν
T ] ≤ lim infT→∞
1
TE[log VT ]
= L = limT→∞
1
Tlog VT ≤ lim inf
T→∞
1
Tlog V πν
T ,
from which the result follows.
5.3. Comparison of functionally controlled and functionally generated portfolios.Recall that, as we observed from the expression in (5.5), the log-optimal portfolio π belongsto the class of functionally controlled portfolios, provided that the drift characteristic λ—asintroduced in the model (5.1)—is sufficiently smooth. In fact, the log-optimal portfolio πis known to be even a (classical) functionally generated portfolio if λ can be written in thegradient form
λ(x) = ∇ logG(x) =∇G(x)
G(x), x ∈ ∆d
+,
for some differentiable function G: ∆d+ → R+; see [CSW19, Proposition 4.7].
Considering again the stochastic model in (5.1), we shall show in this section that thelog-optimal portfolio may genuinely not be a functionally generated portfolio, but still afunctionally controlled one, in cases when λ is not of the above gradient type. We willthen illustrate numerically that the difference between the true log-optimal portfolio andan approximate “best” portfolio based on a class of gradient type trading strategies can besubstantial. This demonstrates that such extensions beyond classical functionally generatedportfolios are crucial.
PORTFOLIO THEORY WITH ROUGH PATHS 35
Let us consider a so-called volatility stabilized market model of the form in (5.1), where,for some γ > 0, the diffusion matrix is given by
cij(µ) := γµi(δij − µj), i, j = 1, . . . , d,
where δij is the Kronecker delta, and the drift is given by
c(µ)λ(µ) = Bµ,
where B ∈ Rd×d is defined by Bij := 1+α2 (1 − δijd) for some α > γ − 1. In the context
of stochastic portfolio theory these models were first considered in [FK05]. The conditionα > γ − 1 assures non-attainment of the boundary, as proved in [Cuc19, Proposition 5.7],i.e. the process µ takes values in ∆d
+.We can solve this linear system for λ, and find as general solution
λi(µ) =1 + α
2γµi+ C, i = 1, . . . , d,
for an arbitrary C ∈ R. Note that this is well-defined as µ always stays within the interior ofthe unit simplex ∆d
+ due to the condition α > γ−1. We now define the function fα:Rd+ → Rby
fα(x) :=1 + α
2γ
d∑i=1
log(xi) + Cd∑i=1
xi.(5.14)
Then ∂ifα(x) = (1 + α)/(2γxi) + C for i = 1, . . . , d, so that
λ(x) = ∇fα(x) = ∇ logG(x), x ∈ ∆d+,
where G(x) := exp(fα(x)). Hence, in this volatility stabilized model the log-optimal portfolioπ can be realized as a functionally generated portfolio. It follows from (5.6) that
supπ∈Π
E[log V πT ] =
(1 + α)2
8γ
(E[ ∫ T
0
d∑i=1
1
µisds
]− d2T
).
A generalization of this model is a polynomial model with the same diffusion matrix (forsome fixed γ), but a more general drift matrix B just satisfying Bjj = −
∑i 6=j B
ij and Bij ≥ 0
for i 6= j (see [Cuc19, Definition 4.9]). In this case λ is in general no longer of gradient type.To see this, let d = 3, and
B =
−p q rp −q 00 0 −r
(5.15)
for p, q, r > 0 such that 2 min(p, q, r) − γ ≥ 0, where the latter condition is imposed toguarantee non-attainment of the boundary (see [Cuc19, Propostion 5.7]). We refer also to[Cuc19, Theorem 5.1] for the relation to (NUPBR) and relative arbitrages.
The solution λ of c(x)λ(x) = Bx is now found to be
λ1(x) =1
γ
(r − p+ q
x2
x1+ r
x3
x1
)+ C,
λ2(x) =1
γ
(r − q + p
x1
x2
)+ C,
λ3(x) = C,
36 ALLAN, CUCHIERO, LIU, AND PROMEL
which cannot be realized as a gradient, for instance since ∂λ3
∂x16= ∂λ1
∂x3.
Let us now compare the log-optimal portfolio
(πt)i = µit
(λi(µt) + 1−
d∑j=1
µjtλj(µt)
)with the functionally generated portfolio
(παt ) = µit
(∂if
α(µt) + 1−d∑j=1
µjt∂jfα(µt)
),
with fα as defined in (5.14). We seek the value of α which optimizes
supα
E[log V πα
T ].
By (5.1) and (5.3), we have that
E[log V πα
T ] = E[ ∫ T
0∇>fα(µs)Bµs ds− 1
2
∫ T
0∇>fα(µs)c(µs)∇fα(µs) ds
]=
1 + α
2γE[ ∫ T
0
(1
µ1s
, . . . ,1
µds
)Bµs ds
]− (1 + α)2
8γ
(E[ ∫ T
0
d∑i=1
1
µisds
]− d2T
).
Since this expression is concave in α, we find the optimizer α∗ to be given by
α∗ =2E[ ∫ T
0
(1µ1s, . . . , 1
µds
)Bµs ds
]E[ ∫ T
0
∑di=1
1µis
ds]− d2T
− 1.
Note that if B is the drift matrix of a volatility stabilized market model with parameter α, theright-hand side yields exactly α, and we find the correct log-optimal portfolio. However, whenwe take πα
∗as an approximate portfolio, for instance in the case of B being of the form (5.15),
this leads to Figure 1. There, with the parameters p = 0.15, q = 0.3, r = 0.2, the functions
t 7→ E[log Vt] (blue) and t 7→ E[log V πα∗
t ] (orange) are plotted, where the expected value iscomputed via a Monte Carlo simulation. This shows a significantly better performance ofthe log-optimal portfolio and, thus, illustrates a clear benefit from going beyond functionallygenerated portfolios in stochastic portfolio theory.
Remark 5.6. Note that for almost all ω ∈ Ω, the log-optimal portfolio (realized by thescenario ω)
πit(ω) = µit(ω)
(λi(µt(ω)) + 1−
d∑j=1
µjt (ω)λj(µt(ω))
)is a functionally controlled portfolio with respect to the rough path µt(ω) in the sense ofExample 4.4. The results in this section illustrate that in the presence of an appropriateprobabilistic structure the asymptotic growth rate can be significantly improved for scenariosoutside a null set, as we anticipated in Remark 4.15.
PORTFOLIO THEORY WITH ROUGH PATHS 37
Appendix A. On the rough path foundation
In this appendix we collect some results regarding rough integration, including its associa-tivity and a Fubini type theorem. While such elementary results are well-known for stochasticIto integration and other classical theories of integration, the presented results seem to benovel in the context of rough path theory and are essential for the model-free portfolio theorydeveloped in the previous sections.
Throughout this section we will consider a general p-rough path X = (X,X)—that is,we will not impose Property (RIE)—and, as usual, we will assume that p, q and r satisfyAssumption 3.2, so that in particular 1 < p/2 ≤ r < p ≤ q <∞.
A.1. Products of controlled paths. As a first step towards the associativity of roughintegration, we show that the product of two controlled paths is again a controlled path; see[FH20, Corollary 7.4] for a similar result in a Holder-rough path setting.
Lemma A.1. Let X ∈ Cp-var([0, T ];Rd). The product operator Π, given by
VqX([0, T ];Rd)× VqX([0, T ];Rd)→ VqX([0, T ];Rd),((F, F ′), (G,G′)) 7→ (FG, (FG)′),
where
(FG)i := F iGi and ((FG)′)ij := (F ′)ijGi + F i(G′)ij
for every 1 ≤ i, j ≤ d, is a continuous bilinear map, and comes with the estimate
(A.1) ‖(F, F ′)(G,G′)‖VqX≤ C(1 + ‖X‖p)2‖F, F ′‖VqX‖G,G′‖VqX ,
where the constant C depends on p, q, r and the dimension d. We call Π((F, F ′), (G,G′)) theproduct of (F, F ′) and (G,G′), which we sometimes denote simply by FG.
Proof. It is clear from its definition that Π is a bilinear map. Suppose (F, F ′), (G,G′) ∈ VqX .For all 1 ≤ i, j ≤ d and (s, t) ∈ ∆[0,T ], we have
‖(FG)′‖q . ‖F ′‖q‖G‖∞+‖F ′‖∞‖G‖q+‖F‖q‖G′‖∞+‖F‖∞‖G′‖q. (‖F‖∞+‖F‖q+‖F ′‖∞+‖F ′‖q)(‖G‖∞+‖G‖q+‖G′‖∞+‖G′‖q)(A.2)
. (1 + ‖X‖p)2‖F, F ′‖VqX‖G,G′‖VqX .
To identify the remainder RFG, we compute
(FG)is,t = F is,tGis + F isG
is,t + F is,tG
is,t
=
( d∑j=1
(F ′)ijs Xjs,t + (RF )is,t
)Gis + F is
( d∑j=1
(G′)ijs Xjs,t + (RG)is,t
)+ F is,tG
is,t
=
d∑j=1
((F ′)ijs G
is + F is(G
′)ijs
)Xjs,t + (RF )is,tG
is + F is(R
G)is,t + F is,tGis,t
=
d∑j=1
((FG)′)ijs Xjs,t + (RFG)is,t,
38 ALLAN, CUCHIERO, LIU, AND PROMEL
where (RFG)is,t := (RF )is,tGis + F is(R
G)is,t + F is,tGis,t. Using the fact that 2r ≥ p, we then
estimate
‖RFG‖r . ‖RF ‖r‖G‖∞+‖F‖∞‖RG‖r+‖F‖2r‖G‖2r
. (1 + ‖X‖p)(‖RF ‖r‖G,G′‖VqX+‖F, F ′‖VqX‖R
G‖r)
+ ‖F‖p‖G‖p(A.3)
. (1 + ‖X‖p)2‖F, F ′‖VqX‖G,G′‖VqX .
The estimate (A.1) then follows from (A.2) and (A.3).
A.2. Associativity of rough integration. The following proposition provides an associa-tivity result for rough integration.
Proposition A.2. Let X = (X,X) be a p-rough path and let (Y, Y ′), (F, F ′), (G,G′) ∈ VqX be
controlled paths. Then, the pair (Z,Z ′) := (∫ ·
0 Fu dGu, FG′) ∈ VqX , and we have that∫ ·
0Yu dZu =
∫ ·0YuFu dGu,
where on the left-hand side we have the integral of (Y, Y ′) against (Z,Z ′), and on the right-hand side we have the integral of (Y F, (Y F )′) against (G,G′), each defined in the sense ofLemma 2.6.
Proof. The fact that (Z,Z ′) ∈ VpX follows from the estimate in (2.5) combined with the
relation Gs,t = G′sXs,t +RGs,t. It also follows from (2.5) that the function H∫F dG, defined by
Zs,t =
∫ t
sFu dGu = FsGs,t + F ′sG
′sXs,t +H
∫F dG
s,t
for (s, t) ∈ ∆T , has finite p-variation for some p < 1, and we can thus conclude that
lim|P|→0
∑[s,t]∈P |H
∫F dG
s,t |= 0. We similarly obtain∫ t
sYu dZu = YsZs,t + Y ′sZ
′sXs,t +H
∫Y dZ
s,t ,∫ t
sYuFu dGu = YsFsGs,t + (Y F )′sG
′sXs,t +H
∫Y F dG
s,t ,
with
lim|P|→0
∑[s,t]∈P
|H∫Y dZ
s,t |= lim|P|→0
∑[s,t]∈P
|H∫Y F dG
s,t |= 0.
Noting that (Y F )′ = Y F ′ + Y ′F , we then calculate∫ t
sYu dZu = YsZs,t + Y ′sZ
′sXs,t +H
∫Y dZ
s,t
= Ys
(FsGs,t + F ′sG
′sXs,t +H
∫F dG
s,t
)+ Y ′sFsG
′sXs,t +H
∫Y dZ
s,t
= YsFsGs,t + (YsF′s + Y ′sFs)G
′sXs,t + YsH
∫F dG
s,t +H∫Y dZ
s,t
=
∫ t
sYuFu dGu −H
∫Y F dG
s,t + YsH∫F dG
s,t +H∫Y dZ
s,t .
Taking lim|P|→0
∑[s,t]∈P on both sides, we obtain
∫ T0 Yu dZu =
∫ T0 YuFu dGu.
PORTFOLIO THEORY WITH ROUGH PATHS 39
Remark A.3. Denoting the integration operator by •, the result of Proposition A.2 maybe expressed formally as Y • (F • G) = (Y F ) • G. We therefore refer to this result as theassociativity of rough integration.
A.3. The canonical rough path lift of a controlled path. Given a p-rough path X =(X,X) and a controlled path (Z,Z ′) ∈ VqX , one can use Lemma 2.6 to enhance Z in a canonicalway to a p-rough path Z = (Z,Z), where Z is defined by
(A.4) Zs,t :=
∫ t
sZu dZu − ZsZs,t, for (s, t) ∈ ∆[0,T ],
with the integral defined as in (2.4). Indeed, we observe the following.
Lemma A.4. Let X = (X,X) be a p-rough path and (Z,Z ′) ∈ VqX be a controlled path.Then, Z = (Z,Z), as defined in (A.4), is a p-rough path. Moreover, if (Y, Y ′) ∈ VqZ , then(Y, Y ′Z ′) ∈ VqX and ∫ T
0Yu dZu =
∫ T
0Yu dZu,
where on the left-hand side we have the rough integral of (Y, Y ′) against Z, and on the right-hand side we have the integral of (Y, Y ′Z ′) against (Z,Z ′) as defined in (2.4).
Proof. That Z = (Z,Z) is a p-rough path follows immediately from Lemma 2.6. That(Y, Y ′Z ′) ∈ VqX can be shown in a straightforward manner using the definition of controlledpaths. Arguing similarly as in the proof of Proposition A.2 and using the same notation, wecalculate, for (s, t) ∈ ∆[0,T ],∫ t
sYu dZu = YsZs,t + Y ′sZs,t +H
∫Y dZ
s,t
= YsZs,t + Y ′s
(Z ′sZ
′sXs,t +H
∫Z dZ
s,t
)+H
∫Y dZ
s,t
=
∫ t
sYu dZu −H
∫Y dZ
s,t + Y ′sH∫Z dZ
s,t +H∫Y dZ
s,t .
Taking lim|P|→0
∑[s,t]∈P on both sides, we obtain
∫ T0 Yu dZu =
∫ T0 Yu dZu.
A.4. The exponential of a rough path. Based on the bracket of a rough path (recall Defi-nition 2.9), one can introduce the rough exponential analogously to the stochastic exponentialof Ito calculus.
Lemma A.5. For a one-dimensional p-rough path X = (X,X) (so that in particular X isreal-valued) such that X0 = 0, we introduce the rough exponential by
Vt := exp(Xt −
1
2[X]t
), t ∈ [0, T ].
Then V is the unique controlled path in VpX satisfying the linear rough differential equation
(A.5) Vt = 1 +
∫ t
0Vu dXu, t ∈ [0, T ],
with Gubinelli derivative V ′ = V .
40 ALLAN, CUCHIERO, LIU, AND PROMEL
Proof. Applying the Ito formula of Proposition 2.10 with Y = X − 12 [X], Y ′ = 1 and f =
exp, we observe that the Young integrals cancel, so that V does indeed satisfy (A.5). Theuniqueness of solutions to (A.5) follows from the stability of rough integration, provided inthis setting by [FZ18, Lemma 3.4].
A.5. A Fubini type theorem for rough integration. In this subsection we provide aFubini type theorem for Bochner and rough integrals. A result of this type is mentioned in aHolder-rough path setting in [FH20, Exercise 4.10].
Theorem A.6. Let X = (X,X) be a p-rough path, let A be a measurable subset of VqX , andlet ν be a probability measure on A. If
∫A‖K,K
′‖VqX dν <∞, then∫ T
0
∫AKu dν dXu =
∫A
∫ T
0Ku dXu dν.
Proof. Due to∫A‖K,K
′‖VqX dν < ∞, the controlled path∫A(K,K ′) dν ∈ VqX exists as a
well-defined Bochner integral. For s < t, we have∫A
∫ t
sKu dXu dν −
∫AKs dν Xs,t −
∫AK ′s dν Xs,t =
∫A
(∫ t
sKu dXu −KsXs,t −K ′sXs,t
)dν
and, by the estimate in (2.3),
(A.6)
∣∣∣∣ ∫ t
sKu dXu −KsXs,t −K ′sXs,t
∣∣∣∣ ≤ C(‖RK‖r,[s,t]‖X‖p,[s,t]+‖K ′‖q,[s,t]‖X‖ p2,[s,t]).
Since 1/r+ 1/p > 1, there exists a p > p such that 1/r+ 1/p = 1. By Holder’s inequality, forany partition P of [0, T ], we have∫
A
∑[s,t]∈P
‖RK‖r,[s,t]‖X‖p,[s,t] dν ≤∫A
( ∑[s,t]∈P
‖RK‖rr,[s,t]) 1r( ∑
[s,t]∈P
‖X‖pp,[s,t]
) 1p
dν
≤∫A‖RK‖r,[0,T ] dν ‖X‖
pp
p,[0,T ]
(max
[s,t]∈P‖X‖
p−pp
p,[s,t]
).
Since∫A‖R
K‖r,[0,T ] dν ≤∫A‖K,K
′‖VqX dν < ∞, and since (s, t) 7→ ‖X‖p,[s,t] is uniformly
continuous, we deduce, treating the second term on the right-hand side of (A.6) similarly,that
lim|P|→0
∑[s,t]∈P
∫A
(∫ t
sKu dXu −KsXs,t −K ′sXs,t
)dν = 0.
Thus, we obtain∫A
∫ T
0Ku dXu dν = lim
|P|→0
∑[s,t]∈P
∫A
∫ t
sKu dXu dν
= lim|P|→0
∑[s,t]∈P
∫AKs dν Xs,t +
∫AK ′s dν Xs,t =
∫ T
0
∫AKu dν dXu.
PORTFOLIO THEORY WITH ROUGH PATHS 41
Appendix B. Rough path theory assuming Property (RIE)
In this section we provide additional results concerning rough path theory assuming Prop-erty (RIE), and, in particular, we give a proof of Theorem 2.12. As usual we adopt Assump-tion 3.2.
B.1. On the bracket of a rough path. We begin with some properties of the bracket of arough path, introduced in Definition 2.9.
Lemma B.1. Let X = (X,X) be a p-rough path and let (K,K ′) ∈ VqX . Recall from Propo-
sition A.2 that (Z,Z ′) := (∫ ·
0 Ku dXu,K) ∈ VqX . Let Z = (Z,Z) be the canonical rough pathlift of Z, as defined in (A.4), so that in particular the bracket [Z] of Z exists. Then
[Z] =
∫ ·0
(Ku ⊗Ku) d[X]u,
where the right-hand side is defined as a Young integral.
Proof. Since [X] has finite p/2-variation, the integral∫ T
0(Ku ⊗Ku) d[X]u = lim
|P|→0
∑[s,t]∈P
(Ks ⊗Ks)[X]s,t
exists as a Young integral. In the following we shall abuse notation slightly by writingHs,t = o(|t− s|) whenever a function H satisfies lim|P|→0
∑[s,t]∈P |Hs,t|= 0. We have
[Z]s,t = Zs,t ⊗ Zs,t − 2Sym(Zs,t)= (KsXs,t +K ′sXs,t)⊗ (KsXs,t +K ′sXs,t)− 2(Z ′s ⊗ Z ′s)Sym(Xs,t) + o(|t− s|)= (KsXs,t)⊗ (KsXs,t)− 2(Ks ⊗Ks)Sym(Xs,t) + o(|t− s|)= (Ks ⊗Ks)[X]s,t + o(|t− s|).
Taking lim|P|→0
∑[s,t]∈P on both sides, we obtain [Z]T =
∫ T0 (Ku ⊗Ku) d[X]u.
Proposition B.2. Suppose that S ∈ C([0, T ];Rd) satisfies (RIE) with respect to p and(Pn)n∈N. Let S = (S,S) be the associated rough path as defined in (2.7). Let (K,K ′) ∈ VqSand (Z,Z ′) = (
∫ ·0 Ku dSu,K) ∈ VqS. Let Z = (Z,Z) be the canonical rough path lift of Z as
defined in (A.4), so that in particular the bracket [Z] of Z exists. Then the following hold:
(i) The bracket [Z] has finite total variation, and is given by
[Z]t = limn→∞
Nn−1∑k=0
Ztnk∧t,tnk+1∧t ⊗ Ztnk∧t,tnk+1∧t, t ∈ [0, T ].
(ii) Let Γ be a continuous path of finite p/2-variation. Then the path Y := Z + Γ admitsa canonical rough path lift Y = (Y,Y), such that
(B.1) [Y]t = [Z]t = limn→∞
Nn−1∑k=0
Ytnk∧t,tnk+1∧t ⊗ Ytnk∧t,tnk+1∧t, t ∈ [0, T ].
42 ALLAN, CUCHIERO, LIU, AND PROMEL
Proof. (i) Since, by Lemma 2.11, [S] has finite variation, it follows from Lemma B.1 that thesame is true of [Z]. By the estimate in (2.3), we know that Zs,t = KsSs,t +K ′sSs,t +Hs,t forsome H satisfying lim|P|→0
∑[s,t]∈P |Hs,t|= 0. It follows that
limn→∞
Nn−1∑k=0
Ztnk∧t,tnk+1∧t ⊗ Ztnk∧t,tnk+1∧t = lim
n→∞
Nn−1∑k=0
(Ktnk∧tStnk∧t,t
nk+1∧t)⊗ (Ktnk∧tSt
nk∧t,t
nk+1∧t)
=
∫ t
0(Ku ⊗Ku) d[S]u = [Z]t.
(ii) Since Γ has finite p/2-variation, the Young integrals∫ ts Zs,u ⊗ dΓu,
∫ ts Γs,u ⊗ dZu and∫ t
s Γs,u ⊗ dΓu are well-defined, and the function Y, defined by
Ys,t = Zs,t +
∫ t
sZs,u ⊗ dΓu +
∫ t
sΓs,u ⊗ dZu +
∫ t
sΓs,u ⊗ dΓu,
also has finite p/2-variation. It follows that Y = (Y,Y) is a p-rough path. The equality[Y]t = [Z]t follows easily from the integration by parts formula for Young integrals. Thesecond equality in (B.1) follows by a similar argument to the one in the proof of part (i).
B.2. Proof – the rough integral as a limit of Riemann sums.
Proof of Theorem 2.12. Let (Y, Y ′) ∈ VqS . Recalling the Ito formula for rough paths (Propo-sition 2.10), it follows from the associativity of Young and rough integrals (recall Proposi-tion A.2) that ∫ t
0Yu df(S)u =
∫ t
0YuDf(Su) dSu +
1
2
∫ t
0YuD2f(Su) d[S]u.
By [PP16, Theorem 4.19], we have∫ t
0YuDf(Su) dSu = lim
n→∞
Nn−1∑k=0
YtnkDf(Stnk )Stnk∧t,tnk+1∧t,
the convergence being uniform in t ∈ [0, T ]. By [FH20, Lemma 5.11], we have the pointwiseconvergence
(B.2) limn→∞
Nn−1∑k=0
YtnkD2f(Stnk )S⊗2tnk∧t,t
nk+1∧t
=
∫ t
0YuD2f(Su) d[S]u.
Recalling Polya’s theorem (see e.g. [Rao62]), which asserts that pointwise convergence ofdistribution functions on R to a continuous limit implies the uniformity of this convergence, wesee from the proof of [FH20, Lemma 5.11] that the convergence in (B.2) also holds uniformlyfor t ∈ [0, T ]. Thus, we obtain
(B.3)
∫ t
0Yu df(S)u = lim
n→∞
Nn−1∑k=0
(YtnkDf(Stnk )Stnk∧t,t
nk+1∧t +
1
2YtnkD2f(Stnk )S⊗2
tnk∧t,tnk+1∧t
),
PORTFOLIO THEORY WITH ROUGH PATHS 43
where the convergence is uniform in t ∈ [0, T ]. For every n and k, we have, by Taylorexpansion,
Ytnk f(S)tnk∧t,tnk+1∧t
= YtnkDf(Stnk )Stnk∧t,tnk+1∧t +
1
2YtnkD2f(Stnk )S⊗2
tnk∧t,tnk+1∧t
+ YtnkRtnk∧t,t
nk+1∧t,
(B.4)
where
Ru,v :=
∫ 1
0
∫ 1
0
(D2f(Su + r1r2Su,v)−D2f(Su)
)S⊗2u,v r1 dr2 dr1.
Since f ∈ Cp+ε, we have that |Ru,v|. |Su,v|p+ε, from which we see that R has finite p/(p+ε)-variation. Since p/(p+ ε) < 1, it follows that
limn→∞
Nn−1∑k=0
YtnkRtnk∧t,t
nk+1∧t = 0,
where the convergence is uniform in t ∈ [0, T ]. Thus, taking limn→∞∑Nn−1
k=0 in (B.4) andsubstituting into (B.3), we deduce the result.
References
[ABI19] Eduardo Abi Jaber, Bruno Bouchard, and Camille Illand, Stochastic invariance of closed sets withnon-Lipschitz coefficients, Stochastic Processes and their Applications 129 (2019), no. 5, 1726–1748.
[Ana20] Anna Ananova, Rough differential equations with path-dependent coefficients, PreprintarXiv:2001.10688 (2020).
[CF10] Rama Cont and David-Antoine Fournie, Change of variable formulas for non-anticipative functionalson path space, J. Funct. Anal. 259 (2010), no. 4, 1043–1072.
[CG98] V. V. Chistyakov and O. E. Galkin, On maps of bounded p-variation with p > 1, Positivity 2 (1998),no. 1, 19–45.
[CO96] T. M. Cover and E. Ordentlich, Universal portfolios with side information, IEEE Transactions onInformation Theory 42 (1996), no. 2, 348–363.
[Cov91] Thomas M. Cover, Universal portfolios, Math. Finance 1 (1991), no. 1, 1–29.[CP19] Rama Cont and Nicolas Perkowski, Pathwise integration and change of variable formulas for contin-
uous paths with arbitrary regularity, Trans. Amer. Math. Soc. Ser. B 6 (2019), 161–186.[CSW19] Christa Cuchiero, Walter Schachermayer, and Ting-Kam Leonard Wong, Cover’s universal portfolio,
stochastic portfolio theory, and the numeraire portfolio, Mathematical Finance 29 (2019), no. 3, 773–803.
[Cuc19] Christa Cuchiero, Polynomial processes in stochastic portfolio theory, Stochastic processes and theirapplications 129 (2019), no. 5, 1829–1872.
[CW21] Steven Campbell and Ting-Kam Leonard Wong, Functional portfolio optimization in stochastic port-folio theory, Preprint arXiv:2103.10925 (2021).
[CZ93] Vijay K. Chopra and William T. Ziemba, The effect of errors in means, variances, and covarianceson optimal portfolio choice, The Journal of Portfolio Management 19 (1993), no. 2, 6–11.
[dF40] Bruno de Finetti, Il problema dei ”Pieni”. Giornale dell’ Istituto Italiano degli Attuari 11, 1–88;translation (Barone, L. (2006)): The problem of full-risk insurances. Chapter I. The risk within asingle accounting period., Journal of Investment Management 4(3) (1940), 19–43.
[DGU07] Victor DeMiguel, Lorenzo Garlappi, and Raman Uppal, Optimal Versus Naive Diversification: HowInefficient is the 1/N Portfolio Strategy?, The Review of Financial Studies 22 (2007), no. 5, 1915–1953.
[Dup19] Bruno Dupire, Functional Ito calculus, Quant. Finance 19 (2019), no. 5, 721–729.[Ebe16] Andreas Eberle, Markov processes, Lecture Notes at University of Bonn (2016).[Fer99] Robert Fernholz, Portfolio Generating Functions, pp. 344–367, River Edge, NJ. World Scientific,
1999.
44 ALLAN, CUCHIERO, LIU, AND PROMEL
[Fer01] , Equity portfolios generated by functions of ranked market weights, Finance Stoch. 5 (2001),no. 4, 469–486.
[Fer02] E. Robert Fernholz, Stochastic portfolio theory, Springer, 2002.[FH20] Peter K. Friz and Martin Hairer, A course on rough paths with an introduction to regularity structures,
2nd ed., Universitext, Springer, Cham, 2020.[FK05] Robert Fernholz and Ioannis Karatzas, Relative arbitrage in volatility-stabilized markets, Annals of
Finance 1 (2005), no. 2, 149–177.[FK11] Daniel Fernholz and Ioannis Karatzas, Optimal arbitrage under model uncertainty, The Annals of
Applied Probability 21 (2011), no. 6, 2191 – 2225.[FL16] Damir Filipovic and Martin Larsson, Polynomial diffusions and applications in finance, Finance and
Stochastics 20 (2016), no. 4, 931–972.[Fol81] H. Follmer, Calcul d’Ito sans probabilites, Seminar on Probability, XV (Univ. Strasbourg, Strasbourg,
1979/1980), Lecture Notes in Math., vol. 850, Springer, Berlin, 1981, pp. 143–150.[FV10] Peter K. Friz and Nicolas B. Victoir, Multidimensional stochastic processes as rough paths. Theory
and applications, Cambridge University Press, 2010.[FZ18] Peter K. Friz and Huilin Zhang, Differential equations driven by rough paths with jumps, J. Differ-
ential Equations 264 (2018), no. 10, 6226–6301.[GR13] Massimo Guidolin and Francesca Rinaldi, Ambiguity in asset pricing and portfolio choice: a review
of the literature, Theory and Decision 74 (2013), no. 2, 183–217.[HKT01] Friedrich Hubalek, Irene Klein, and Josef Teichmann, A general proof of the dybvig-ingersoll-ross-
theorem: Long forward rates can never fall, Preprint arXiv:math/0112230 (2001).[HS10] Hardy Hulley and Martin Schweizer, M6—on minimal market models and minimal martingale mea-
sures, Contemporary quantitative finance, Springer, Berlin, 2010, pp. 35–51.[IL22] David Itkin and Martin Larsson, Robust asymptotic growth in stochastic portfolio theory under long-
only constraints, Math. Finance 32 (2022), no. 1, 114–171.[IP15] Peter Imkeller and Nicolas Perkowski, The existence of dominating local martingale measures, Fi-
nance Stoch. 19 (2015), no. 4, 685–717.[Jam92] Farshid Jamshidian, Asymptotically Optimal Portolios, Mathematical Finance 2 (1992), no. 2, 131–
150.[KF09] Ioannis Karatzas and Robert Fernholz, Stochastic portfolio theory: an overview, Special Volume:
Mathematical Modeling and Numerical Methods in Finance (Alain Bensoussan and Qiang Zhang,eds.), Handbook of Numerical Analysis, vol. 15, Elsevier, 2009, pp. 89–167.
[KK07] Ioannis Karatzas and Constantinos Kardaras, The numeraire portfolio in semimartingale financialmodels, Finance Stoch. 11 (2007), no. 4, 447–493.
[KK20] Ioannis Karatzas and Donghan Kim, Trading strategies generated pathwise by functions of marketweights, Finance Stoch. 24 (2020), no. 2, 423–463.
[KR12] Constantinos Kardaras and Scott Robertson, Robust maximization of asymptotic growth, The Annalsof Applied Probability 22 (2012), no. 4, 1576–1610.
[KR17] Ioannis Karatzas and Johannes Ruf, Trading strategies generated by Lyapunov functions, FinanceStoch. 21 (2017), no. 3, 753–787.
[KR21] Constantinos Kardaras and Scott Robertson, Ergodic robust maximization of asymptotic growth,Ann. Appl. Probab. 31 (2021), no. 4, 1787–1819.
[LCL07] Terry J. Lyons, Michael Caruana, and Thierry Levy, Differential equations driven by rough paths,Lecture Notes in Mathematics, vol. 1908, Springer, Berlin, 2007.
[Lej12] Antoine Lejay, Global solutions to rough differential equations with unbounded vector fields, Seminairede Probabilites XLIV, Lecture Notes in Math., vol. 2046, Springer, Heidelberg, 2012, pp. 215–246.MR 2953350
[LH14] Bin Li and Steven C. H. Hoi, Online Portfolio Selection: A Survey, ACM Comput. Surv. 46 (2014),no. 3.
[LQ02] Terry Lyons and Zhongmin Qian, System control and rough paths, Oxford University Press, 2002.[Mar59] Harry Markowitz, Portfolio selection: efficient diversification of investments, no. 16, New Haven:
Yale University Press, c1959, 1970 printing., 1959.[MPS11] Eberhard Mayerhofer, Oliver Pfaffel, and Robert Stelzer, On strong solutions for positive definite
jump diffusions, Stochastic Processes and their Applications 121 (2011), no. 9, 2072–2086.
PORTFOLIO THEORY WITH ROUGH PATHS 45
[PP16] Nicolas Perkowski and David J. Promel, Pathwise stochastic integrals for model free finance, Bernoulli22 (2016), no. 4, 2486–2520.
[Pro04] Philip E. Protter, Stochastic integration and differential equations, 2nd ed., Springer, 2004.[PW07] Georg Pflug and David Wozabal, Ambiguity in portfolio selection, Quantitative Finance 7 (2007),
no. 4, 435–442.[PW16] Soumik Pal and Ting-Kam Leonard Wong, The geometry of relative arbitrage, Math. Financ. Econ.
10 (2016), no. 3, 263–293.[Rao62] R. Rao, Relations between weak and uniform convergence of measures with applications, The Annals
of Mathematical Statistics 33 (1962).[RX19] Johannes Ruf and Kangjianan Xie, Generalised Lyapunov functions and functionally generated trad-
ing strategies, Appl. Math. Finance 26 (2019), no. 4, 293–327.[SSV18] Alexander Schied, Leo Speiser, and Iryna Voloshchenko, Model-free portfolio theory and its functional
master formula, SIAM J. Financial Math. 9 (2018), no. 3, 1074–1101.[Str14] Winslow Strong, Generalizations of functionally generated portfolios with applications to statistical
arbitrage, SIAM J. Financial Math. 5 (2014), no. 1, 472–492.[SV16] Alexander Schied and Iryna Voloshchenko, Pathwise no-arbitrage in a class of delta hedging strategies,
Probab. Uncertain. Quant. Risk 1 (2016), Paper No. 3, 25.[Won15] Ting-Kam Leonard Wong, Universal portfolios in stochastic portfolio theory, Preprint
arXiv:1510.02808 (2015).
Andrew L. Allan, ETH Zurich, SwitzerlandEmail address: [email protected]
Christa Cuchiero, University of Vienna, AustriaEmail address: [email protected]
Chong Liu, University of Oxford, United KingdomEmail address: [email protected]
David J. Promel, University of Mannheim, GermanyEmail address: [email protected]